INFORMATION TO USERS This reproduction was made from a copy of a m anuscript sent to us for publication and microfilming. While the m ost advanced technology h as been used to pho­ tograph and reproduce this m anuscript, the quality of the reproduction is heavily dependent upon the quality of the material submitted. Pages In any m anuscript may have indistinct print. In all cases the best available copy has been filmed. The following explanation of techniques is provided to help clarify notations which may appear on this reproduction. 1. M anuscripts may not always be complete. When it is not possible to obtain missing pages, a note appears to indicate this. 2. When copyrighted materials are removed from the m anuscript, a note ap­ pears to indicate this. 3. Oversize materials (maps, drawings, and charts) are photographed by sec­ tioning the original, beginning at the upper left hand corner and continu­ ing from left to right in equal sections with small overlaps. Each oversize page is also filmed as one exposure and is available, for an additional charge, as a standard 35mm slide or in black and white paper format. * 4. Most photographs reproduce acceptably on positive microfilm or micro­ fiche b u t lack clarity on xerographic copies made from the microfilm. Fbr an additional charge, all photographs are available in black and w hite standard 35mm slide format.* *For more information about black and white slides or enlarged paper reproductions, please contact the Dissertations Customer Services Department TU T-TV/T-T Dissertati°n 1 V 1 1 Information S e r v ic e University Microfilms International A Bell & Howell Information Company 300 N. Zeeb Road, Ann Arbor, Michigan 48106 8700480 H u d so n , W illiam D e n n is CLASSIFICATION OF CONIFEROUS FOREST COVER TYPES USING LANDSAT MULTISPECTRAL SCANNER DIGITAL DATA Michigan State University University Microfilms International 300 N. Zeeb Road, Ann Arbor, Ml 48106 PH.D. 1986 PLEASE NOTE: In all c a se s this material h as been filmed In the best possible w ay from th e available copy. Problems encountered with this docum ent have been identified here with a ch eck mark V 1. Glossy photographs or p ag es 2. Colored illustrations, paper or p rin t. 3. Photographs with dark background A. Illustrations are p o o r c o p y _______ 5. P ages with black marks, not original c o p y ______ 6. Print shows through as there is text on both sides of p a g e _______ 7. Indistinct, broken or small print on several p ag es. 8. Print exceeds m argin req u irem en ts______ 9. Tightly bound co p y with print lost in s p in e _______ ^ y y y 10. Computer printout p ages with indistinct print_______ 11. P ag e(s)____________ lacking when material received, an d not available from school o r author. 12. P ag e(s)____________ seem to b e missing in numbering only as text follows. 13. Two pages n u m b e re d 14. Curling and wrinkled p a g e s ______ 15. Dissertation co n tain s p ages with print at a slant, filmed a s received 16. Other____________________________________________________________________ _____ . Text follows. ^ University Microfilms International CLASSIFICATION OF CONIFEROUS FOREST COVER TYPES USING LANDSAT MULTISPECTRAL SCANNER DIGITAL DATA By William Dennis Hudson A DISSERTATION in partial Submitted to Michigan State University fulfillment of the requirements for the degree of DOCTOR OF PHILOSOPHY Department of Forestry 1986 ABSTRACT CLASSIFICATION OF CONIFEROUS FOREST COVER TYPES USING LANDSAT MULTISPECTRAL SCANNER DIGITAL DATA By William Dennis Hudson The major objective of this study was to evaluate Landsat multispectral scanner digital data for classifying and mapping coniferous forests. All analysis was conducted over two test sites in the north central Lower Peninsula of Michigan using a winter, Existing snow-covered, Landsat scene. "standard" digital analysis techniques tested included unsupervised clustering, and maximum likelihood algorithms. minimum d i s t a n c e - t o - m e a n s , In addition, level slicing and reducing the maximum allowable cluster radius were implemented with the clustering technique. A spectral response curve model was developed from the analysis of the multispectral reflectance patterns exhibited by the coniferous cover types and the background features. Predicted brightness values were used to develop a linearcombination classifier. The classifier uses the value obtained by subtracting BV5 from BV6 in conjunction with the absolute band 6 brightness value (BV6, B V 5 - B V 6 ). Digital classification techniques had accuracies similar to those obtained from the visual interpretation of computer enhanced spring imagery over both the Crawford County test site, (73.8 versus 72.7 percent) and the Wexford County test site classification (84.0 versus 84.8 percent). Supervised techniques were g ene ra ll y more accurate and executed faster than unsupervised clustering techniques. The linear-co mbi nation classifier (BV6, BV6-BV5) was the most accurate and one of the fastest digital classifiers t e s t e d , regardless of test site. Overall classification accuracy was significantly different between the two test sites. The least accurate classification for Wexford County exceeded the most accurate classification for Crawford County. With respect to acc uracy and relative efficiency, li nea r-combination classsifier recommended digital (B V 6 , BV6-BV5) the is the technique under the conditions tested. To Jan ice ii ACKNOWLE DGM ENTS I am especia lly appreciative of the efforts of Dr. Ramm, my major advisor/ Carl for both his personal and professional support throughout this study. I would also like to express my appreciation to the other members of my committee; Dr. James B. Hart, Dr. Gary A. Simmons, and Dr. David P. Lusch for their man y valuable suggestions. This work was completed at the Center for Remote Sensing whose staff contributed important technical assistance. A special thanks to Tamsyn Mihalus for her patience and special efforts. This research was supported by a National Aeronautics and Space Administration grant, Michigan State University, NASA NGL 23-004-083, Center iii for Remote Sensing. to TABLE OF CONTENTS Page LIST OF T A B L E S ............................................ vi LIST OF F I G U R E S .......................................... x CHAPTER ........................................ I. INTRODUCTION A. Forest Management Data Needs ................. B. Remo tel y Sensed D a t a .......................... C. O b j e c t i v e s ...................................... 1 1 3 5 II. FOREST CLASS IF ICA TION FROM LANDSAT DATA . . . . . III. STUDY DESIGN AND TECHNIQUES ....................... A. Scene/Test Site S e l e c t i o n ................... B. Digital Classification Procedures .......... 1. I n t r o d u c t i o n ............................. 2. Un supervised Classification ............ 3. Supervised Techniques ................... C. Acc uracy Assessment Procedures ............... 1. Forest Cover Type M a p s ................. 2. Aerial Photograp hy .............. D. Algorithm Development ........................ 1. Characteristic Response Curves . . . . 2. Response Curve Modeling ................. 3. Line ar-Combination Classifier .......... E. Kappa Coefficient of Agreement .............. 7 20 20 27 27 29 30 31 31 32 35 35 39 42 43 IV. RESULTS AND D I S C U S S I O N ............................ 48 A. Classification Error Analysis .............. 48 1. I n t r o d u c t i o n ............................. 48 2. Un su pervised Clustering . . . . . . . . 52 3. Supervised Classification ............... 92 107 4. Linear-Combination Classifier .......... 5. Bou ndary E r r o r s ............................. 122 B. Effects of Accuracy Assessment Procedures . 128 C. Relative Effici enc y of Classifiers .......... 130 V. SUMMARY AND C O N C L U S I O N S ............................... 137 iv APPENDICES APPENDIX A. THE LANDSAT S Y S T E M ............................... 148 B. SNOWFALL D A T A ................................. 186 C. DIGITAL CLA SSIFICATION OF LANDSAT MULTISPECTRAL SCANNER DATA ................ 191 D. COM PUTER PROGRAM TO CALCULATE A KAPPA STATISTIC AND VARIANCE FROM AN N x N .......................... 238 CON TINGENCY TABLE BIB LIOGRAPHY ............................................... V 242 LIST OP TABLES Table Page 1. Summary of Landsat MSS training sites ............... 36 2. Confusion table format ............................... 50 3. Clustering statistics* default parameters, Wexford County test s i t e ............................... 53 4. Two -w ay cross tabulation, default cluster, Wexford County test s i t e ............................... 55 5. Landsat classification performance, default clusters, Wexford County test site ................ 58 6. Clustering statistics, default parameters, Crawford County test s i t e ...............................61 7. Two-w ay cross tabulation, default clusters, Crawford Cou nt y test s i t e ............................... 62 8. Landsat classification performance, default clusters, Crawford County test site ................ 63 9. Clustering statistics, smaller cluster radius, Wexford County test s i t e ..................... 67 10. Two- way cross tabulation, smaller cluster radius, Wexford County test s i t e ..................... 69 11. Landsat classification performance, smaller cluster radius, Wexford County test site ......... 70 12. Clustering statistics, smaller cluster radius, Crawford County test s i t e ............................... 72 13. Two- way cross tabulation, smaller cluster radius, Crawford County test site ................... 73 Landsat cla ssification performance, smaller cluster radius, Crawford County test site ......... 75 14. vi Clustering statistics/ default parameters on level sliced scene/ Wexford County test site . . . 77 Two-way cross tabulation/ default parameters on level sliced scene/ Wexford County test s i t e ................................................... 79 Landsat classification performance/ default parameters on level sliced scene/ Wexford County test site . ............................................ 80 Clustering statistics/ default parameters on level sliced scene/ Crawford County test site 81 . . . Two-way cross tabulation/ default parameters on level sliced scene/ Crawford Cou nt y test s i t e ................................................... 82 Landsat classification performance/ default parameters on level sliced scene/ Crawford County test site ..................................... 83 Clustering statistics/ smaller cluster radius on level sliced scene/ Wexford County test site . . . 85 Two -wa y cross tabulation/ smaller cluster radius on level sliced scene/ Wexford County test site ............................................... 86 Landsat classification performance/ smaller cluster radius on level sliced scene/ Wexford County test site ..................................... 87 Clustering statistics/ smaller cluster radius on level sliced scene/ Crawford County test s i t e ................................................... 89 Two-way cross tabulation/ smaller cluster radius on level sliced scene/ Crawford County test site ............................................... 90 Landsat classification performance/ smaller cluster radius on level sliced scene/ Crawford County test site ............................ 91 Training site signatures/ Wexford County test site ............................................... 93 Training site classification/ minimum distanceto-means classifier/ Wexford Cou nty test site . . . 94 • • VII Training site classification/ maximum likelihood classifier/ Wexford County test site ............ 95 Two-way cross tabulation/ minimum distance-tomeans classifier/ Wexford County test site . . . 96 Landsat classification performance/ minimum distanc e-t o-means classifier/ Wexford County test site ............................................ 97 Two-w ay cross tabulation/ maximum likelihood classifier/ Wexford County test site ............ 99 Landsat cla ssification performance/ maximum likelihood classifier/ Wexford Cou nty test site . 100 Training site signatures/ Crawford County test s i t e ................................................. 102 Training site classification/ minimum distanceto-means classifier/ Crawford County test site 103 Training site classification/ ma xim um likelihood classifier/ Crawford County test site . . . . . . 104 Two-way cross tabulation/ minimum distanceto-means classifier/ Crawford County test site 105 Landsat classification performance/ minimum dis tan ce-to-means classifier/ Crawford County test site ............................................ 106 Two -wa y cross tabulation/ maximum likelihood classifier/ Crawford County test site ............ 108 Landsat classification performance/ maximum likelihood classifier/ Crawford County test s i t e ................................................. 109 Two -wa y cross tabulation/ B V 6 - B V 5 / Wexford County test site ................................... 110 Landsat classification performance/ B V 6 - B V 5 / Wexford County test site .......................... 112 Two-way cross tabulation/ B V 6 - B V 5 / Crawford County test site ................................... 113 Landsat classification performance/ B V 6 - B V 5 / Crawford County test site .......................... 114 viii Two- way cross tabulation, Wexford County test site B V6, BV6-BV5, ............................ 116 Landsat classification performance, B V 6 , BV6-BV5, Wexford County test site ................... 121 Two-w ay cross tabulation, BV6, B V 6 - B V 5 , Crawford County test site ............................ 123 Landsat cla ssification performance, B V 6 - B V 5 , Crawford Cou nt y test site BV6, ................ 127 Number of omission and commission errors determ ine d from maps and aerial photography, ............................ Wexford County test site 128 Number of omi ssion and commiss ion errors determined from maps and aerial photography, Crawford Coun ty test site ............................ 129 Summary of Landsat classification performance, Wexford Cou nty test site ............................ 131 Summary of Landsat classific ation performance, Crawford County test site ............................ 132 Comparison of classific ati on algorithm attributes ............................................ 135 Comparison of Landsat coverage with aircraft coverage ............................................... 178 Snowfall and snow on the ground, Digital values, for bands 4, February 1979 . . 187 5, 6, 7, and 8 . . . . 220 Mean brightn ess values, by band, for 27 clusters ............................................... 222 ERDAS training sample statistics listing, Landsat filename: T1 .......................................... 231 Signature list for file: 234 T1 ....................... Landsat classification continge ncy table ix ......... 235 LIST OF FIGURES Figure 1. Page Landsat black-a nd- white band 5 scene E-30358-15471 22 2. Location of scene E-30358-15471 and the two test s i t e s .................................................... 23 3. Test site 1 (western p o r t i o n ) , west central Wexford County (NHAP 81-31-90) ..................... 25 Test site 1 (eastern portion), west central Wexford County (NHAP 81-233-239) ................... 26 Test site 2, northeastern Crawford County and southeastern Otsego County (NHAP 81-3-215) . . . . 28 4. 5. 6. Forest cover type map, Wexford County test s i t e ...................................................... 33 7. Forest cover type map, Crawford County test s i t e ...................................................... 34 8. Spectral response curves (plot of mean br ightness values from selected training sites). for red pine . . . . . 37 9. Spectral response curve model 41 10. Plot of mean br ig htn ess values per band for 27 c l u s t e r s ................................................. 54 11. Forest error map, default clusters, Wexford County test s i t e ........................................ 59 12. Wexford County test site (portion), A-omitted forest, B-comroitted forest .......................... 60 13. Forest error map, default clusters, Crawford County test s i t e ........................................ 65 14. Crawford County test site (portion), A-omission of lightly stocked jack p i n e .......................... 66 x 15. Crawford County test site (portion), B-commission of hardwood s t a n d ................................... 16. Forest error map, smaller cluster radius, Wexford County test 3 i t e ............................... 71 17. Forest error map, smaller cluster radius, Crawford 76 County test s i t e ............................... A-l. Landsat ob ser vat ory configuration A-2. Inclination of Landsat orbit to maintain sun synchronous orbit ............................... A-3. A-4. A-5. Sun synchronous Landsat orbit ............. Landsat orbital tracks for one day of c o v e r a g e ........................................ 154 .. Landsat nominal A— 10. RBV scanning p a t t e r n .......................... A-ll. Spectral response, A-12. Multispectral scanning arrangement A-13. Ground scan pattern for a single MSS detector 157 ......... RBV three-camera system 158 159 . .. ............. A - 1 4 . Landsat bla ck-and-white band 4, scene E-30556-15460 155 155 A-9. A— 16. 151 Relationship of actual scene centers to nominal center points for Landsat scenes acquired on a repetitive basis ................................... image center locations 151 153 Variations in azimuth of solar illumination with season and l a t i t u d e ...................... A-7. A-15. . .Sun illumination relationships for 45° north latitude . . . . . . . . . Orbital coverage characteristics of Landsat A-8. 150 .................. A-6. 66 161 162 . 162 164 Landsat black-and-white band 5, scene E-30556-15460 164 Landsat black-a nd- white band 6, scene E-30556-15460 165 xi A— 17• Landsat black-and-white band 7, scene E-30556-15460 A-18. 165 Re lationship between voltage and digital count for a hypothetical scan l i n e .............. 166 A— 19. Formation of the MSS picture e l e m e n t ........... 168 A-20. Scanning pattern of the RBVs on Landsat 3 . . . 169 A-21. Format of a fully processed Landsat 3 RBV scene showing the location of subscenes A , B, C , and D . .............. A — 22. Landsat RBV subscene A , E-30052-15461 A-23. Landsat black-an d-whit e band 8, E-30052-15462 A-24. A-25. A-26. A-27. A-28. A-29. A-30. B— 1. B-2. B-3. C-l. 170 172 . 173 Landsat image coverage, latitude and longitude lines drawn for an area in C o l o r a d o ............ 174 Correspondence between a Landsat image and computer compatible tapes ....................... 175 Landsat annotation block explanation Landsat false-color composite, E-21487-15352 ........... 176 February 1979, 181 Landsat false-color composite, April 1976, E-2443— 15421 ....................................... 182 Landsat false-color composite, E-30556-15460 183 September 1979, Landsat false-color composite, acquired under low sun elevation conditions, November 1978, E - 2 1379-15300 Location of weather stations 184 .................... 188 Percentage of years during which a 6-inch or greater snow depth occurred ..................... 189 Average number of days per season with accumulated snow depth on the ground of 6 inches or m o r e ..................................... 190 Landsat image (E-30358-15471), black-and-white band 5 ............................................... 192 C-2. Correspondence between Landsat MSS image gray scale and digital brightness values from CTT . . 194 Two-dimensional plot of training site pixel v a l u e s ............................................... 196 Geometric interpretation of the minimum distanc e-t o-mean s classifier ..................... 198 C-5. Geometric derivation of distance ................. 199 C-6. Geometr ic interpretation of the parallelepiped classifier/ with rectangular ................................. decision regions 202 Geometric interpretation of the parallelepiped classifier/ with parallelogram decision regions ................................... 203 Classification based upon probability density f u n c t i o n s ........................................... 205 Points of equal pr obability as defined by the maxi mum likelihood classifier ................... 207 C-3. C-4. C-7. C-8. C-9. C-10. Unsupervised clustering .......................... C-ll. Plot of mean br ightness values for 27 clusters . C-12. Flow diagram illustrating the major components for Landsat classification ....................... 218 221 226 C-13. Transparent grid for determining line and column numbers from a Landsat image ............ 227 C-14. Portion of the statistics and historgram listings available from program BUILDH ......... 228 C-15. Outlining an area of known cover t y p e .......... 230 C-16. His togram of training site d a t a ................. 230 C-17. Flow diag ram of signature manipulation p r o g r a m s ............................................ 233 C-18. Classified image from the MAXCLAS program . . . C-19. Individual pixel assignment to classes (27) from an unsupervised clustering ................ xiii 237 237 CHAPTER I INTRODUCTION A. Forest Management Data Needs Modern progressive forest resource management consists of three interrelated activities: assessmenti 1) resource inventory and 2) development of management strategies and alternatives/ and 3) implementation of specific achieve desired goals. programs to The inventory and assessment phase is a ne ce ssa ry pre-requisite for intelligent decision-making required in the latter stages. Traditional timber- orientated inventory systems contain data on forest location/ forest cover type/ conditions/ productivity/ forest utilization/ recreation/ acreage/ and ownership. including wildlife/ dispers ed recreation/ and minerals/ volumes/ has assumed stand The diversi ty of developed timber products/ and water increasing significance in management policy and has thus intensified forest resource data needs. In response to detailed/ needs in decision-making/ current/ and repetitive data foresters have developed and adopted a vari ety of innovative inventory tools. Paramount among these techniques are statistical sampling schemes/ remotely sensed data/ and computer inventory systems. Although these techniques may be applied independently/ they are often utilized simul ta neo usly in an integrated resource inventory program. An important element in the implementa­ tion of forest inventories has been the di chot omy of increasing data needs in a time of restricted budgetary allocations. Several current inventories/ Michigan/ using examples taken from may be cited to illustrate current capabilities and limitations. The Forest Management Division/ Department of Natural Resources Michigan (D N R ), is conducting compartment and stand ex aminations as part of an operations inventory (Rose, 1978). This inventory is designed to collect multiple stand variables and to make specific management treatment rec ommendations for use in day-to-day operations on the forest. Although it represents one of the most intensive efforts to assess forest resources, require ten years to complete and is confined forests. it will to state owned Similar systems are being used on national forest holdings and to varying degrees by forest industries. The Michigan Resource Inventory Act (Public Act 204) provides for the preparation of a state-wide current use inventory (Michigan Inventory Advisory Committee, Maps depicting land cover/use will be compiled 1982). from aerial photography and this map-based information will be entered into a computerized data base. Although this program will m a p the entire state, most forest lands will only be classified at a very general level pine, lowland conifers). (e.g., central hardwoods, The status of Michigan's timber volumes is periodically assessed (as are all states) Analysis (formerly called Resources Evaluation and earlier known as Forest Surv ey ). 1 U.S. Forest Service, through Forest Inventory and This inventory, conducted by the measures and evaluates timber conditions as well as the supply and drain situation. However, years, the inventory is conducted only every 10 to 15 is not site-specific, and is statistically reliable only on a multi- county basis. Current forest resource data gathering efforts are inadequate to meet immediate needs. managers must, therefore, Forest resource investigate and adopt new data collection techniques which are both efficient and costeffective. B. Rem otely Sensed Data Aerial photographs have long been recognized as a useful, if not critical, tool in the forest inventory. The favorable vantage point and the synoptic view of aerial photography are especi all y suitable in extensive forest situations where limited ground observations may be the rule. Aerial photographs are currently considered integral part of ma ny forestry activities. to be an Foresters utilize aerial photography to inventory timber stands, ^■Michigan's most recent forest inventory (1980) is reported on in the following references: Raile and Smith, 1983: Spencer, 1982 and 1983: Hahn, 1982: Jakes, 1982: and Smith, 1982. determine site classes, and to monitor forests for insect and disease infestations. are, However valuable aerial photos they do have limitations and are used to complement, improve, or reduce field work, The past two decades, not to replace it. the "space age," have spawned a number of new remote sensing capabilities. This period has been one of significant advances in sensor technology, platform development (high-altitude aircraft and spacecraft) and information extraction techniques. Thermal scanners are now utilized on an operational basis for monitoring forest fires, earth orbiting satellites are used for change detection studies and automated interpretation techniques are presently being developed. While several orbiting systems have been developed and tested (e.g. Skylab and the Heat Capacity Happing Mission), the Landsat series has been the primary vehicle of the National Aeronautics and Space Administ rati on' s observation program. (NASA) land Although still considered by many to be in the developmental stage, the Landsat program has generated considerable interest from both the scientific community and resource managers. Since the launch of ERTS-1 Satellite-1, Ju ly 1972, orbit. (Earth Resources Technology latter renamed Landsat (Land Satellite) -1) in four additional satellites have been placed in Currently, Landsat-4 and Landsat-5 are operational. The large aerial coverage (about 13,225 square miles per scene)/ the 9 or 18-day repetitive acquisition cycle, and the multispectral nature of the data all may be exploited for forestry applications. Researchers and resource ma nagement agencies have analyzed Landsat multispectral scanner (MSS) data both visually (image interpretation) and by utilizing computer compatible tapes (automatic digital locations, cat egory definitions, processing). Study and classification techniques are nearl y as varied as the number of studies undertaken. Likewise, accuracies have varied among studies, from under 10 percent (rarely reported in the literature) to over 90 percent. In spite of numerous research studies and several federally funded technology transfer programs, operational use of Landsat data for forest resource assessments is minimal, being virtually nonexistent in Michigan. This study was conducted to thoroughly test certain existing digital analysis techniques, algorithms where necessary, to devel op new classification and to recommend an inventory system which uses Landsat technology for forest resource analysis. C. Objectives The major objective of this study is to evaluate Landsat MSS digital data for classifying and mapping coniferous forests. Several existing digital analysis techniques will be analyzed to determine the extent to which these techniques could meet the information needs of forest managers. In addition, a new classification algorit hm is proposed which utilizes a linear-combination of Landsat bands. Finally, the use of Landsat technology in an inventory system will be addressed. Specific research tasks designed to accomplish the overall objective 1. include: Evaluate several (i.e. "standard" digital analysis techniques algorithms available on the ERDAS m i cro ­ computer): to-means, unsupervised clustering, min imum distanceand maximum likelihood. Determine the accuracy of coniferous forest type maps developed by these procedures. 2. Devel op and test a linear-combination classifier from a model of characteristic response curves. Determine the ability of the classifier to identify and map coniferous forest types. 3. Analyze the relative effici ency of the classifiers, including a comparative analysis of the magnitude and sources of errors. 4. Develop a strategy for a coniferous forest inventory which uses the Landsat data. Based on the performance of the various classifiers, specific recommendations will be made in a comprehensive inventory system. for using Landsat data CHAPTER II FOREST CLA SS IFICATIO N FROM LANDSAT DATA A v a ri ety of sensor configurations have been utilized in the Landsat system to date. The primary sensor on each of these satellites is the Multispectral Scanner (MSS). Detectors emp loyed on the Landsat MSS system record the reflected radiation of the earth's surface from an area of 79 x 79 meters. The amount of radiation that reaches the detector is measured in four spectral bands. Two of these spectral bands are within the visible light portion of the electromagnetic spectrum/ green and red, and two are from the reflected infrared portion. as com pute r-compatible tapes Landsat data are available (CCT) in a digital format or as photo-like products derived from the digital data. Landsat system is described in more detail The in Appendix A. The Landsat program has generated and funded numerous investigations to test, document and apply MSS data to forest resource evaluations. of studies undertaken, Unfortunately, many do not provide adequate documentation of the procedures algorithms) of the hundreds (i.e. classification used and/or an assessment of accuracy obtained. Individual classification techniques based upon forest conditions throughout the U.S. and in the Lake States in particular are summarized. The U.S. Forest Service has conducted two comprehensive evaluations of Landsat data for forest and rangeland inventories. Georgia/ The first study analyzed data from sites in Colorado/ and South Dakota land could be distinguished ( H e l l e r / 1975). Forest from non-forest land with 90 to 95 percent accuracy with either photointerp ret ation or computer assisted techniques. The identification of individual cover types could not be made with acceptable accuracy using either method. A later study tested computer processing techniques for inventorying forest and grassland resources within ten broad ecological communities found in the U.S. (Mazade/ 1981); Rocky Mountain Conifer Forest/ Northern Hardwood Forest/ Juniper Forest/ Boreal Forest/ Northern Conifer Forest/ Oak-Pine Forest/ Grassland/ Hardwood Forest. Pinyon- Southeastern Pine Forest/ Pacific Coast Forest/ and Central Results indicated that automated processing of Landsat MSS data could di stinguish classes of softwood/ hardwood/ better accuracy. categories (e.g. grassland/ and water with 70 percent or The identification of more specific pine from spruce/ maple from oak) was not ge ner ally obtainable. The Canadian Forestry Service has investigated the use of single date and multidate com binations of Landsat MSS data for mapping forest resources in Canada 1979/ and 1981). (Kalensky et a l . An area in northern Saskatchewan was classified into several broad resource classes: reg eneration and brush/ land. A two-date recent burn/ (August-May) softwood/ and unproductive supervised/ maximum forest likelihood classification yielded the best results, accuracy of 90 percent. single date (August) Supervised classifications using a and three-date {August-May-March) composite were somewhat less accurate. rectangular parallelepiped classifier, An unsupervised, single-date classification yielded the least accuracy. forests (coniferous forest, land) with an Mapping of deciduous forest, in southeastern Ontario was also and non-forest investigated. Classification accuracy using a maximum likelihood classifier of single-date scenes varied significantly as a function of the date of acquisition. obtained for a September scene lowest accuracy scene. (67 percent) The increase (one 2-date, Highest accuracy was (81 percent) was obtained for an October in accuracy for multidate combinations two 3-date, and one 4-date combination) marginal and ranged from one to two percent. high correlation whereas the was Because of the found between bands 4-5 and between 6-7, classifications were performed on a two band data set, band 5 (red) and band 7 (second near-infrared). High accuracies were obtained with two two-band data sets using either a single date percent), (82 percent) indicating or a multidate combination (84 that these two bands contain sufficient data for forest classification. A study to determine the ability to classify forest cover types in Pen nsylvania tested data from September and January using both an unsupervised clustering and a 10 supervised, min im um d i s t a n c e - t o - m e a n s c l a s s i f i e r al. Con if er ou s forest types could be d i f f er en ti ated 1974). from decidu ou s or wi nt er data, possible. types by both cl as si fier s using eit her summer but Merging discrimination. final maps: hardwood, further di ffer en t i a t i o n was not the two data sets improved forest type Six cover types were diffe re nt ia ted on the hardwoods, fields, shaded hardwoods, and water. conifers, hemlock- Fort y points were randomly located on the ma p and then located was (Borden et found to not agree with the map, in the field. One point "for an a cc uracy es ti mat e based on the grou nd check method of 98 percent." Several additional studies have likewise been limited to d i s c ri mi na ti ng ge ne ra l i z e d forest classes only. An analysis of Landsat-1 MSS data over the Feath er River W at er shed in Cal if or nia evalua te d both manual and automated techniques including (Krumpe et a l . 1973). four timber cla ss es hardwoods, overall Seven resource classes, (conifers, and c o ni fe r plantations) ac c u r a c y of 79 percent. were chaparral, identified with an The product ion of forest cover type maps in Maine used supervised cl as si fi catio n techniques wood, to map the following categories: hardwood, and Bryant, water, 1983). open, Overall and other softwood, (Bryant et al. mixed 1980 agreement wa s 63 percent whereas d i f f e renc es betw ee n acr ea ge s obt aine d by Landsat compared an existin g inven to ry were under five percent. data over No r t h Carolina was used to de te r m i n e to Landsat MSS if a Southern 11 pine forest could be stratified into crown closure classes (Williams, 1976). Classification of hardwoods, subdivided into three crown closure classes, and clearcuts was based upon spectral training areas within a two date registered scene. 70 percent, pine regeneration, signatures from (February and August) Overall classification acc uracy was only regrouping the pine stratification into a single class yielded an accuracy of 90 percent. Although com monly recommended (Hazade, 19B1), for large area mapping detailed forest classifications have been unattainable using unsupervised clustering techniaues. An evaluation of two clustering algorithms in northern Idaho indicated that neither algorithm could separate types (Werth, 1981). generalized classes (meadow), Classification accuracy for 4 (coniferous and water) algorithms. forest cover forest, cut-over, was 60 and 67 percent grass for the two A study in western Washington attempted to separate nine species/size/stocking groups using unsuper­ vised clustering. The classification agreed in only 56 percent of the plots whereas more generalized classes (composed of clear-cut, forest classes) agreed established plantations, and older in 87 percent of the sample plots. A land cover map emphasizing forest types for the entire state of California has been prepared using unsupervised clustering classes, techniques (Newland et a l . 1980). Sixteen including six forest classes (conifer, conifer- 12 hardwood, hardwood, h a r d w o o d - w o o d l a n d ), h a r d w o o d - c o n i f e r , c o n i f e r - w o o d l a n d , and were identified and mapped. The study did not include an assessment of accuracy. Guided clustering techniques have been utilized by a number of investigators and, conditions, under certain specific site have suc cessfully identified forest cover types. This technique combines the character is tic s of supervised and unsupervised methods to develop a large number of low variance classes (Fox and Mayer, 1979). Training defined as in a normal supervised classification, fields are and a minimum distance clustering algorithm is used to create spectrally distinct classes from the original fields. Unsupervised clustering training is also used on the data set to generate a second set of spectrally distinct classes. The two class lists are then merged. are spectrally similar are pooled; Pairs of classes which classes that are similar to two or more classes are deleted (indicating an illdefined c l a s s ) ; and classes that are spectrally unique are retained. Each class is then assigned to a resource category and class statistics are utilized to classify the entire scene, such as by a maximum likelihood classifier. Salazar (1982) utilized guided clustering to derive fuel type maps as a means of providing basic pre-attack planning for fire management in northwestern California. Ten fuel type categories were identified and mapped from which an overall mean pro bability of correct classification 13 (0.71) was computed. Two variations of guided clustering were also evaluated in southwestern Colorado Hoffer, 1980). conifer, grass, (Nelson and Gen eralized forest categories and barren) were mapped with average accuracies of between 76 and 88 percent. (hardwood, Although two additional studies indicate that individual species are identifiable using guided clustering techniques, neither study provided an assessment of acc uracy (Khorram and Katibah, 1981 and Mor rissey and Ambrosia, 1982). Coniferous forest cover types have been successfully discriminated using guided clustering techniques. has been partially attributed to environmental Success factors, espec ial ly the tendency of many forest cover types to be topographically distributed in mountainous terrain. Coniferous tree species within Crater Lake National Park, Oregon were classified classification into seven classes with an average accuracy of 89 percent. 2 size classes, Four conifer groups, and 2 den sit y classes were Shasta-Trinity National Forest, identified on the California with 86 percent accuracy, whereas the classification of just the tree species was 91 percent accurate. The use of ancil lar y or collateral data to improve Landsat classifications has been investigated by several researchers. A v a riety of techniques have been developed for combining continuo us image data with spatial categorical collateral data (Strahler et a l . 1980). Two basic 14 approaches have been utilized. The first treats the collateral data as if they were additional bands or channels and uses the collateral data to estimate prior probabilities to be used in a maxi mum likelihood classification. A second approach utilizes ancillary data to modify labels or provide more detailed descriptions of classes as a post classif ica­ tion technique. Digital aspect, terrain data, including elevation, slope, and have been utilized in conjunction with Landsat MSS data to differentiate between species and stands with similar spectral characte ristics but with different habitat requirements. A topographic distribution model was developed to statistically characterize the distribution of forest cover types in the San Juan Mountains in southwestern Colorado (Fleming and Hoffer, types were classified 1979). Twelve forest cover from the merged data sets with accuracies ranging from 54 to 80 percent. Accuracies of 39 to 59 percent were obtained utilizing spectral data only. Similar approaches were used to map forest fuel classes in Montana with an accuracy of 68 percent compared to a classification of 52 percent without the use of terrain data (Shasby et al. 1981 and Burgan and Shasby, channel data set 1984). A 15- (12 spectral channels from ^hree-date Landsat data and 3 terrain channels) was used to di ff eren ­ tiate ten forest cover types in northern California (Strahler et a l . 1978). An interactive parallelepiped 15 classification with terrain data was between 82 and 85 percent accurate whereas a maximum likelihood decision rule using probabilities based upon field sampling was between 71 and 77 percent accurate. A regional Eldorado National Forest, California was generated from Landsat MSS data and digital ecological model. type map of the terrain data using an Forest cover types and stand size classes were mapped with an overall accuracy of 84 percent. additional studies reported Two that species differentiation and height/density stratification was possible with the addition of terrain data, but neither provided any estimates of classification accuracy (Strahler et a l . 1979 and Woodcock et a l . 1980). Ancillary data has also been utilized in a post­ classification mode by several researchers. resource categories Fifteen including 9 forest categories were mapped with a guided clustering technique in Humbolt County, California with an overall classification accuracy of 82 percent (Fox and Mayer, 1981). Seven climatological/ vegetational regions were defined and then mapped onto the Landsat classification. The addition of the ecological zones permitted more detailed descriptions of resource categories such that the number of unique categories increased from 15 to 40. A label relaxation technique has been developed by Richards et a l . (1982) wherein Landsat classification labels are modified by probabilities from 16 ancillary (elevation) data. An analysis of the spruce-fir category versus other categories in the Colorado Rockies indicated that classification accuracy could be increased from 68 to 80 percent. One last method eliminates small disjointed sets by combining the areas and relabeling them (Kan, 1976). Mixed softwood/hardwood stands, which were not mappable using standard techniques, were identified using merger criteria for mixed stands with 32 percent accuracy. Results from investigations conducted States have, regions. for the most part, paralleled those from other Landsat MSS data were utilized by Mead and Meyer (1977a and 1977b) be mapped in the Lake to determine which forest categories could in Itasca County, categories, Minnesota. Eleven land cover including three forest classes lowland conifer, and mixed (upland conifer, forest) were correctly classified 43 and 58 percent over two test sites. A study conducted by Hoffer et a l . (1978) compared estimates of forest acreages obtained from Landsat MSS data with those obtained by the Forest Survey. The percent of gross forest area, regardless of location, was compared on a county-by-county basis in four states. coefficients between each state were: Missouri (0.69), Correlation the two sets of forest acreages for Michigan (0.951), Wisconsin and Ne w York (0.90). (0.974), A more detailed comparison of the data in Michigan indicated that relatively low and high estimates for the four Survey Units "balanced 17 o u t ” so that there was a difference of only 1.95 percent between the two methods. A study conducted under a National Aeronautics and Space Administration ERTS-1 Project Grant analyzed both visual and computer assisted 1974). interpretations Results (for forestry tasks) indicated that: hectares) (Myers et a l . from this research 1) large forested tracts (larger than 30 could be delineated and mapped by manual interpretation of Landsat transparencies and 2) well stocked stands were classified by computer analysis with an accuracy of 85 percent. A study to evaluate the accuracy and costs of mapping small forestlands in southwestern Michigan compared two image dates and two image products of each scene et al. 1981). (Karteris A band 5 positive and a false-color composite of a winter (February) composite of a fall scene, and a diazo-enhanced color (September) scene were interpreted. Separate forest/non-forest maps and area estimates were derived from visual interpretation of the four Landsat images. The overall accuracies ranged from 74 to 98 percent, and were higher for the winter scene than the fall scene. Mapping accurac y was highest for the winter false- color composite, scene improved whereas the diazo enhancement of the fall the mapping accuracy over the fall color composite. false- 18 Another study utilized Landsat MSS imagery to determine the location and acreage of forestland in a 30 county site (Hudson and Kittleson, 1978). An electronic scanning de nsi tometer was utilized to perform density slicing of the Landsat images (Hudson, independently, determined forest acreages by performing a band 5 (forest) Two interpreters, - band 7 (water) acreage estimates, prepared 1981). density slice. operating Forest compared with a forest distribution map from aerial photography, averaged 4.7 percent low (ranging from a low of 8.7 percent to an overestimate of 8.5 percent). A study to evaluate the use of Landsat computer enhanced imagery for mapping coniferous forest types has been completed A computer enhanced (Franklin et al. false-color composite, using visual spring 1983). (April) scene was classified interpretation procedures. The accuracy achieved by two interpreters was compared and summarized contingency tables. 85 to 73 percent, in Overall classification accuracies were whereas ability accuracies ranged individual species interpret- from a low of 32 percent for mixed pine stands to 95 percent for jack pine plantations. Utilizing supervised pattern recognition Roller and Visser (1980) in Lake County, attempted to map forest cover types Michigan from June MSS data. from 22 map categories stocking classes) technigues, Signatures (including seven species groups and 2 were used to classify forest cover in the county. An analysis of 195 random samples indicated that species/stocking classes were identified with an overall accuracy of 46 percent. In summary/ it can be concluded that most cla ssifi­ cations of Landsat MSS data have been limited to generalized forest categories deciduous/ (usually some combination of coniferous/ and mixed categories). reported success Those studies which have in identifying forest cover types have g ene rally dealt with coniferous dominated forests in the western U.S. Since many of these forest cover types are topographically distributed/ terrain data has freauently been merged with the Landsat data prior to classification. CHAPTER III STUDY DESIGN AND TECHNIQUES A. Scene/Test Site Selection The climate/ physiography/ and soils of the northern Lower Peninsula of Michigan combine to create a unique region/ as compared to the southern Lower Peninsula or the Upper Peninsula. One of the most distinctive landscapes is the high outwash plains near the center part of the region/ dominated by stands of pines and oaks. flat/ sandy plains — These extensive/ where the large white pine Michigan formerly grew — forests of support much of the nearly 7 million acres of forest found in this part of the state. Hardwoods are more extensive than softwoods with aspen-birch/ a sub-climax type/ the most common forest association throughout the region. The northern hardwood type is most prevalent in the northwestern counties. Lowland hardwoods are primarily restricted to riparian sites. The softwoods/ or coniferous forest types/ ap proximately 22 percent of the region. these conifers are pines; subtype/ accounts for more jack pine/ Three-fifths of the most important pine than half the acreage. occurs as a rel ati vely pure type in a broad belt central northeastern area/ central area. occup y Jack pine from the south and west to the southwest Significant areas are also found in the east and west central counties. 20 21 The swamp conifer types — lesser areas of black spruce, tamarack — land. predominately cedar with balsam fir-white spruce, and comprise less than 9 percent of the forest These swamp conifers occur most frequently as small patches of a few acres on wetlands. The exceptions are in the northeastern counties where these species are more frequent and exist as larger stands. Since the major objective of the study was to classify and map coniferous was chosen forests, for analysis. forestlands in Michigan a winter (snow-covered) scene Previous research on delineating indicated that a snow-covered scene provides a high degree of discrimination between forested and non-forested areas (Karteris, 1981). false color composites also indicated (color) An evaluation of that sharp tonal contrasts existed among several coniferous forest types. The scene chosen for analysis (E - 3 0 3 5 8 - 1 5 4 7 1 ) was obtained by Landsat 3 on February 26, 1979 is referenced as path 23 - row 29 and north-central Lower Peninsula acquisition, (Figure 1). is centered (Figure 2). It in the On the date of there was an average of 58.4 cm of snow on the ground as reported from 17 weather stations, 38.1 to 93.4 cm (NOAA, specific snowfall data. 1979). ranging from Appendix B includes more Almost all of this snow had accumulated by mid February. No new snowfall was recorded just prior to the Landsat overpass. As a result, virtually 22 Figure 1. Landsat black-and-white band 5 scene E-30358 15471. 23 Figure 2. Location of scene E-30358-15471 and the two test sites. 24 all non-forest cover types, exhibited the spectral hardwood including inland lakes, response of snow. forests were leafless, Although the their extensive mass of trunks and branches provided a sub stantially different reflector compared to the underlying snowpack. ferous forests provided The coni­ the only green-foliage reflectances in the entire scene. Two test sites {Figure 2) in the northern Lower Peninsula were chosen to be representative of areas supporting large acreages of conifers. The first test site (Figures 3 and 4) was located in west central Wexford County (T. 22 and 23 N., P . 12 W.) square miles. and enc ompasses ap pr oxi mately 30 It is underlain by stratified sand and gravel outwash deposits confined to a broad valley (i.e., train). time, Most of the area was cleared a valley for agriculture at one but later was abandoned as unsuitable for sustained crop production. The present forest covers more than 37 percent of the area and is primarily pine plantations. for 65 percent of the plantations, in 20 percent of the plantations, mixture of red and jack pine. land is typed as swamp conifer. Red pine accounts jack pine predominates and 15 percent is a Four percent of the forest These stands, composed of scattered northern white- ced ar intermixed with lowland hardwoods, the region. are concentrated along several creeks traversing 25 Figure 3. Test site 1 (western portion)/ west central Wexford County (NHAP 81-31-90). 26 Figure 4. Test site 1 (eastern portion), west central Wexford County (NHAP 81-233-239). 27 The second test site was located in northeastern Crawford County and southeastern Otsego County (T. N . / R.l W.) This area 28 and 29 and encompasses approximately 20 square miles. (Figure 5) and is typical of the is part of an extensive outwash plain "jack pine flats" of the northern Lower Peninsula of Michigan. No evidence of land clearing for agriculture is present and the forests are entirely natural (i.e. there are no plantations). Conifers cover over 70 percent of the area/ remainder being hardwoods/ grass/ or brush. the Pines are located throughout the site except for two large swamps which support lowland conifer species. Relatively pure jack pine is the major cover type throughout the site and represents more than 70 percent of the softwood acreage. Red pine accounts for 4 percent of the acreage/ mixtures 9 percent/ B. pine and swamp conifers 17 percent. Digital Classification Procedures 1. Introduction The "standard" digital analysis techniques evaluated in this study consisted of algorithms currently available on the ERDAS computer. (Earth Resources Data Analysis Systems) micro­ It features both a grid-based geographic information software package and a Landsat digital analysis software package. The ERDAS system is a stand-alone computer system based around a Z-80 Central Processing Unit and supports ASSEMBLER, BASIC, FORTRAN, and PASCAL 28 Figure 5. Test site 2, northeastern Crawford County and southeastern Otsego County (NHAP 81-3-215). 29 programming languages. density/ Other hardware double-sided floppy disk drives/ control,CRT, matrix printer/ drive. includes dual doub le ­ a joystick cursor RGB color monitor, The Landsat software package and tape is a modular interactive system permitting color display of a 240 by 256 pixel, three-band, subscene. false-color image of a user selected The Landsat software site selection, includes supervised visual enhancement, training haze correction, minimum d i s t a n c e - t o - m e a n s , maximum likelihood classification, unsupervised classification. Output and is in the form of a color display and line printer maps. 2. Unsupervised Classification Cluster analysis is a method of sorting data into groups such that the "degree of natural association" within groups and low between groups. tering, a cluster analysis technique, is high Unsupervised clu s­ is utilized to classify Landsat data by creating a series of groups (classes) based on pixel brightness values. Each class will contain a group of pixels with "similar" brightness values and should, therefore, represent a unique cover type. Different cover types, which should have brightness values, "different" would be represented by one or more different cluster classes. The ERDAS unsupervised clu s­ tering algorithm is described and illustrated in Appendix C. A sub-scene from each of the two test sites was utilized to test the effect of varying the values of the 30 input parameters utilized by the clustering routine. conducting numerous repetitive classifications, After the only variable which consistently improved classification performance was the reduction of the maximum allowable cluster radius from seven to three digital counts. were run on the entire Clusters test sites utilizing both default input values and the smaller cluster radius value. In an effort to minimize the effect of a large number of non-forest clusters, level slicing was used to pre- process the Landsat data. An analysis of cluster assignments indicated a maximum brightness value, above which forest pixels were unlikely to occur. per band, All pixels above this threshold were assigned a value of zero and excluded from further consideration. Two clusters, one using default parameters and the other a smaller allowable cluster radius 3. (3), were run on each level sliced test site. Supervised Techniques Supervised classification techniques emp loy user specified training areas to establish quantitative descriptors for each cover type. The resulting training sets are utilized by the computer to classify each pixel based upon a pre-established decision rule. The two decision rules available on the ERDAS system are a minimum distance-to-means and a maximum likelihood classifer (both algorithms are described and illustrated in Appendix C). 31 Training site statistics were obtained from a total of 136 training sites. Each site was characterized from both photo-interpretation of color infrared imagery (1:24,000 scale) and subsequent sites were chosen field verification. The individual to represent as many different stand conditions as possible occurring on the various major landform types. The three pine types included both plantations and natural stands. An iterative procedure (input training sites signatures, evaluate accuracy, classif y test site, delete and/or add new signatures) was used to develop a final set of training site signatures. Subsequently, the Wexford County and Crawford County test sites were classified using the minimum dis tance-to-means and maximum likelihood classifiers. C. Ac curacy Assessment Procedures 1. Forest Cover Type Maps To evaluate the acc uracy of the various algorithms, the Landsat cla ssifications were compared with previously compiled cover type maps. These maps were prepared specifically for this project and were constructed photo interpretation of medium-scale (1:24,000) infrared photography. both U.S.D.A. Additionally, from color Forest Service and Michigan Department of Natural Resources forest cover type maps were consulted in conjunction with ground verification by field crews. sampled in the field. A total of 136 stands were 32 The cover type maps were digitized into polygon format, checked forerrors, re-sampled to match converted into grid the Landsat pixel size. file files, The and re-sampled maps were then rectified to overlay the digitial Landsat files. Eighteen ground (map) control points were matched with their appropriate Landsat pixel location to compute the coefficients of a linear transformation. "rectified" The map thus corresponds with the Landsat scene on a pixel-bypixel (or c e l l - b y - c e l l ) basis. The "digital version" of the Wexford County test site is shown in Figure 6 and that for Crawford County in Figure 7. Having the two files and forest cover types) (Landsat classification results registered meant that cross­ tabulations of the files could be accomplished to create contingency tables. In addition, the registered maps also permitted the production of error maps spatial distribution, and type of error. 2. cover type showing the Aerial Photog rap hy In order to evaluate the effectiveness of the accuracy assessment procedure using cover type maps, a test was conducted using aerial photography as the "ground truth." composite error map was created for each test site by mapping only those pixels which were similarly misclassified by every technique (assuming that these errors, if any, may have been the result of the mapping process). Aerial photographs (Figures 3,4, and 5) were enlarged and A 33 K^ S B S , ^ K‘* ‘ ‘1" *81!'®'$' I " ■”;Vi Mit. ' JfjfM. s'li' '^$9% iS&W3 If'l v»'*^1'' ' iJJ4' «• ; ,13:;:’i *•i* RED PINE SWAMP CONIFERS JACK PINE WATER MIXED PINE Figure 6. Forest cover type map/ Wexford County test site 34 ss s s M m'**" i2*aS2 fS** is■iim i ! m m m 8§ S i 8 S i $ b l0 i.:\\\V** : **«£* -« ■m?*»«£m ssHssm« “• , - im ;»s^sSs^t!^S'en „t.a« ft* ..-i3;-.3«rtUf!:*,nHuwa S S S t n S ^ J l ? if •yH ?„ aa ot ri r nf rr c * i B HS «i n n nB n «murmsr ~uB!»nH ta tnn o u nn» u.ffOBflatisu'i.’T Ba «n tf iu f l i t a « 4 n c 2 B n : « « t a s n o c a ...» B« » »:»« icK Bt »taRunB Hn ue ar aBa SB BBS ac Bt Bi ta nR aa Bn DR jS ib ttBB*...wnKtre w « .rtaoSat B»Jf{ fies* \y\\ S•»tM B tR l!”u». »»«•.m'-«itrtantt*ttnB*R«ai*ao» *S888S8883„.BsnRCB»««aeaa 888S888?...•tfaoBBaaao 8*8*8*8”8 JtfflJ BB B3;t Sn 5s Sn 8w 8l 8i 3s S5 "8888 * .■ o \s .w w w s sTw sU caaaes W ^\«iS% \\ vsv«kBJejnu !’uinanaaiinaiinansTittpnna Hiltl'J B'»' Biia utfs!l3r.t)«U 3't:inne u sttBBaS snssin rtfloSR nuna*-' asiiaaatt asKoaa \lsn-jaw«.***■* aeuuBmtswnBoasis»«*•' brums su cs#a»auam? in \n v iK B vwsosaounBBBBRBcnu u nan **’ W i S i 0 * • ' iSsrW \ k r ? i 8 0 u n a o n n t m u n u s a a a o n r c e c * ijanuiinosHHMaauuau Va'iui;H*in*B:mnuaf.-anBsB33auaB:;f \fiHi;B4BaantneanBetti-ae«BRouc#i «hh \ at nio a o: H« aa t iuo Hco*. ta af BU Rn ao sia R a no 4f ut I«■ awu «m t ; f ; t B n3 c;K B M n ; f t* c \ ti .Bat R 3 a a y f l «at tl !! B; n a a a atf « n •*■-«••*•* \, C ' ij «t am oB nR n r :j u B a n i o a a n c ca !a i f : aB f c . a 3 u a t a » c s B 3 ! M a a o i f n » f'.fiattKUBaBRnKcr«enw:40Boa \ aauausanuMaaaoBaRaviia \HBncanunwoartaanuBaao \nnuusBnflsanuanniiS9sa \ ;B>!ntve»oanuii*siif!aaa V: cJC!iBsnuni»HTiaansita nnt“'inairs'<«»•* tttntiiiattBBB S tuts •H sfl^si^ss’ s«.va. I P I iP f . nn cut* arsuatruyiutwinaft nna «anH4UBKffB»Raott34R"Bsintiit»' 1 “ V. *£■*« ...at-jjaa _flB«t»sKuaflnft88t'.n»«t:» ^iJoantt»r«aaa*-«rtau»««nf5R« »S.>fSK^?a.-j »i?8!85SS888888*888«8S8*88S88», «US 8S 8S 88 88*5888885388«S88SSS8S8 88»8«S ,,„8Sf .„S»-S«iia SKtis«g^i|s! jS. a* »* 8S 88 88 88 S” ’ i88»8 8£8 an': ' £-S8 ,fW* *“8888*888588888585888888 •tiwaacm; JS8S8S8888'8S8S8288"8 UR H 8?* J.iR?.waua s-aa^n ......r-jffis ttitna n«aa o r.aituitafiRJia Ktiifs ...0U BtisnBJiifBrBKaa bukbhu sKit a :iit;iut:B(.‘»atsa»nBu a snit MB.tacwantitfnrwit :vjsn«!iaauan-4RCK«a u a :;»:r t;»»u H-.iaaa^BtBirsijMi aaa •i.‘nJ atve;i:*H«:ta3BatfRj;3rn:vn« "BIt” s r. “,7* 5 a £ /ft#"’ ft *| f fcfc\s a nuti beTd^? m H D k DlMB a an iiit ititBB a HU tin IKHtt v HQ a V Q B P H o asceaatiHs^ iiu mi ,j h ::■t :::•:ii g L* ■“■ 3 b» * If ilMW «5fta tf t nnBttMana f li ii mc ut u. n« ^ui a BROf.fltt QBaRlIirBIlBr.ttrtJlSBflfJ •:fv ‘ »**“ Kf . IJtitfniUfSORIlW s•fttf t .t tn SftUHfloBRBBf.tlBB aet na n ». »,r ~r.BBit ‘Ills “V ff WsiB! tW 'S f! Je ''" n»*»°„ ibb S*?-" gi8 Sf&^i 8fifi Sa8 j * ^sssifig^YXgk .«». J§sisSH.9"a fta,. -*,irS* "!$! $S. *« -»„/« /. ” 'gg*Si!Sg#**'*» \W W ttn u an «pat:itr.aiT|iKijanr.« rkwss- •ia cu « »a a t: a a u a ti •;i;a u aK::-saa :raft vu a u a:f n rn iic f.in a a n a r, ilnuRirs; a« S K r» n * i a»aa uaMsaataviaaaKB IsSKRnKttB RtiSCttBSORriOBRBBOauRBiin • aitanBtinMnosa an b B uf.tiOBt»ttnsB>iBi:KaBHr.»rertttntn KaiJ:iHwa:)ai:!itiH.-:a«'.ty»;faa tRRRnsaiiauv.ftptmrotttfttu a jtBa*«;fl;saji-iur’Btii»Tia»r5'.i3f?a?in:tnMSBiirttjnv.aaH»KHn«.**imj!natiRtiO sKtsanB r titt«?;»»■a.SBH3Hn«ttUJUf»i;«atS75!iKns:'-e«finB!t.".r.»tu!Bf.r.!!ysuB nairrt»sjisjM3Kff£«as.-ffflBuaBflatf«r.3jt3BitMaauj . r4K»:anttBH?ti;Mi:3natjnB ftno ur.H«t;ynr!ftftjtR«BB3RoafitB3«n«j;«ifjiiJiit:r3SiiTBaKacnBi.‘ififtiM:r..u!aK Ertttaasaa.i.iiUfuuoaRaitB.it.a k t.-a.iiicauas ;i:. uyuacaw M m i;‘F . •B»Jrrni'Ruj. r :f*4' u, }:i tnf i,V Lm‘tt ^ § 1 sm RED PINE JACK PINE ggh't:i%s. SWAMP CONIFERS 1* 1 ii iiil WATER MIXED PINE Figure 7. Forest cover type map» Crawford County test site, 35 registered to the error map using the same control identified above. points A systematic sample of one-fourth of the mis-classified pixels was compared with the aerial photography to determine errors of omission/ commission/ and the probable cause of the misclassification. D. Algorithm Development 1. Characteristic Response Curves Digital brightness values (BVs)l were obtained training sites located throughout the Landsat scene. from 136 Most of the coniferous cover types exhibited an exceptionally large range of brightness values individual (Table 1), although training sites were reasonably homogeneous and seldom had a range of over 10 digital counts band. In contrast/ for a single the hardwood training sites not only had a large range of BVs/ but also were much more variable — ranges from 15 to, 25 BVs were not uncommon for a single band. Snow covered areas, on the other hand/ smallest range of BVs; band 5 was saturated but a few cases/ than five digital exhibited the (BV=127) in all while band 6 typically had a ranqe of fewer counts. The mean brightness value for each training site was utilized to plot spectral response curves for individual cover types (Figure 8). While still ^The voltages produced by each MSS detector are converted from an analog signal to a digital form by means of a multipexer. The resulting number/ from 0 to 127/ are called brightness values and are dire ctly related to the amount of solar radiation reflected from the surface of the Earth for a specific wavelength band. Table 1. Cover Type Red Pine Summary of Landsat MSS Training sites. | Range of Mean Brightness Values Number Sample t _______ 1____ — of size | Band 4 Band 5 Band 6 Band 7 Sites (Pixels) I___________ 114.6-58.4 30.0-65.6 26 1 1 .0 - 6 6 . 8 25.7-79.3 1015 1 1 Jack Pine 115.2-60.8 j 12.3-70.4 19.7-74.4 21.2-59.1 30 1341 Mixed Pine 114.8-33.5 1 I j15.4-29.7 12.0-35.4 24.3-4 2.3 26.1-37.0 22 1056 12.7-31.0 21.1-37.8 24.8-34.6 22 1646 33 .9-100 34.1-99.4 30.4-75.5 23 1519 123-127 88.8-96.7 13 1277 39.3-127 21.2-96.7 136 7854 Swamp Conifers Hardwoods 1 1 1 j30.9-83.2 1 1 Snow | 108-122 l ___ All | 14.6-122 126-127 11.0-127 37 100 100 60 Vsluss) TO 70 60 60 60 60 40 4Q 30 30 Di gi tal Count s 60 (Brightness CO 20 10 (a) RED PINE (b) JACK PtNE 4 Spactrst Band 6 6 7 Figure 8 . Spectral response curves (plot of mean brightness values from selected training s i t e s ) . 40 Digital Counit CBrlgMntit Valuta) 38 SHOW 90 10 (c) PINE MIXTURES 4 5 6 50 7 tl Band 70 LEAFLESS HARDWOODS o 40- 40 (Srlghtnafli Vtlut*) 50 30 - Dlgl'il Counli 20 - Cd> S W A M P C O N I F E R S <») N O N - C O N I F E R O U S 5 S picinl Figure 8 Band (c on t' d.). « Bptctral Sand 39 exhibiting a large overall range of brightness v a l u e s , the spectral response curves display a discernable pattern. Coniferous training sites mixtures, (red p i n e , jack pine, and swamp conifers) pine representing highly stocked stands had the lowest spectral responses. The shape of these curves, for example the three lowest red pine curves on Figure 8 a, follows that of "typical" green vegetation: slight decrease from band 4 (green light) light), a substantial (near-IR) a to band 5 (red increase from band 5 to band 6 and a moderate increase from band 6 to band 7. The curves for coniferous vegetation contrast sharply to those for hardwoods and snow which display an increase from band 4 to band 5; a near-constant, relatively small difference between band 5 and band 6 ; and a moderate to substantial decrease from band 6 to band 7. The coniferous training sites exhibiting higher mean spectral responses represent stands with varying degrees of stocking. stocking decreases, As more of the area mapped as coniferous forest is composed of snow, and oc cas ionally hardwoods, which contribute to the spectral response of the stand. Therefore, the brighter coniferous BVs are the result of the combined reflectance of the conifers and snow (or hardwoods) — the lower the stocking level, 2. the higher the brightness. Response Curve Modeling A spectral response curve model was developed from analysis of the m u l t i s p e c t r a l , reflectance patterns 40 exhibited by the coniferous cover types and the background features. Training-site signatures from areas of relatively pure cover types were used as the foundation of the model. The brightness values corresponding to mixtures of various cover types could be predicted by summing the average spectral response for each cover type weighted by its spatial extent in the instantaneous field of view fIFOV): k BVm = s Pi BVi i=l ®vm = predicted brightness value of the mixture of cover types (for a single band) Pi = proportion of cover type (i) BVi = average brightness value for the pure cover type (i) k = in the IPOV (for a single band) number of cover types in mixture As an example, the mean brightness values for snow and pure red pine can be utilized to predict pixel brightness values for various spatial mixtures of these two types representing various stocking levels of red pine. progression of the response curves, model (Figure 9), The general shape and as generated by the is similar to those which were obtained from training-site data (Figure 8 ). The red pine stand represented by curve A (Figure 8 ) had mean brightness values of 40.9, 44.2, 52.8, and 45.2 for bands 4,5,6, and 7 120 " 110too a 80 - > e a 00 - m c 3 oo 00 40 RSO PI NE 20 . 4 0 a 7 S p ectral Band Figure 9. Spectral response curve model for red pine. 42 respectively. According to the model/ (S) and red pine the percent of snow (PR) can be calculated as follows: P s = B V m - BVpR and Pp R = 1 - P s BV S - BVpR Curve A would therefore represent a stand composed of approximately 75 percent red pine and 25 percent snow. actual brightness values predicted from the model particular combination 51.0/ 3. The for this (curve A on Figure 9) are 41.0/ 40.0/ and 46.3. Linear-Co mbi nation Classifier The predicted brightness values a mixture of conifers and background from stands containing features demonstrated that the magnitude of change in reflec tiv ity from band 5 to band 6 provided the most consistent measure ing among the cover types. hypothesis/ using data In order to test this a portion of the Landsat scene was classified from only these two bands. predicted characteristic increase band 5 to band 6 for conifers/ flat response using the compared to the relatively from BV 6 (BV6-BV5) In addition/ (BV6-BV5) Because of the in brightness values from for background features/ subtracting BVS test areas. for dis cr im in at­ the value obtained by was used to classi fy the the test areas were classified data in conjunction with the absolute band 6 brightness values. 43 E. Kappa Coefficient of Agreement Several approaches to evaluating any differences in classification acc uracy have been presented literature. data, If a desired test only includes two sets of Rosenfield and Mell ey (1980) derived 1. from standard statistical Cochran, tests: the significance of the two means (Snedecor and 1967). The Wilcoxon signed rank test for paired samples (Sokal and Rohlf, 3. present three tecnniques A t-test for determining difference between 2. in the 1967). The sign test for paired samples (Snedecor and C o c h r a n , 1967). When more than two sets of data are to be compared, two-way analysis of variance without replication has been proposed as an appropriate test the mean (Rosenfield and Melley, 1980). Since is a proportion and does not satisfy the assumption of normality, an arcsine transformation approximate a normal distribution In addition, (Sokal and Rohlf, if the two categories are as is often the case, is required to 1969). frequently confused, then the proportions corresponding to these categories may not be independent as is assumed use of analysis of variance (Aronoff, In lieu of "standard" statistical multivariate analysis table analysis, techniques, in the 1981). tests, discrete often called contingency have been developed for acc uracy assessment 44 of remotely sensed data (Congalton et al., Congalton and Head 1981 and 1983). These 1981, distributed 1983, techniques were designed for the analysis of discrete data, classification data, 1982, such as which are discrete and mu ltinomially (each cat egory is binomially distributed). The method does not assume that categories are independent and utilizes the entire matrix (confusion table). The problem with using the percent of correct pixels as a measure of agreement is that even a random classifier will pattern of positive 1982). percentages produce a (TurJc, 1979 and Chrisroan, A measure of overall agreement which adjusts agreement which would occur from an independent random classifier has been developed by Cohen statistic, called Kappa, for the R, or KHAT, (1960). This is computed as the difference between the observed agreement of the classifica­ tion (classifier versus reference data as indicated by the diagonals) and the chance (expected) agreement which is defined as the product of the row and column marginals: Kappa = (observed - expected) / (1 - expected) The measure of agreement is thus calculated as (Bishop et al., 1975): R = N r r £ X£i £ i=l i=l N2 - where: r £ i=l (x^. * x.^) (xi* * x.j) r = number of rows (and columns) in the matrix 45 x ii = number of observations in row i and column i (ie. the i ^*1 diagonal element) xj. = the marginal total for row i x.jl = the marginal total for column i N = total number of observations For computational purposes, the following formulae are normally used: R = 0} - Q2 1 9! = ©2 - r Z Xii i=l r ©2 = E x i* * x -i i=l N The kappa statistic will vary from -1, disagreement, to +1, for complete for complete agreement (when K=0, observed agreement equals chance agreement). It can be utilized to compare two error matrices which vary by a single variable. Several formulae for the standard error of the kappa statistic which were previously derived {e.g. Cohen considered 1960 and 1968, in error. and Everitt, 1968) are now The approximate large sample variance of R, derived from the delta method as presented by Rao (1965), al., has been presented by several authors 1969, Bish op et a l ., 1975, and Fleiss, (Fleiss et 1981): 46 ©id-©!) 1 52 c m 2(l-©1 )(2©1© 2-e3 ) + N (l-02 )2 . (l-©2 )3 {l-91 )2 (04-4022 ) + -----------------(l- 0 2 )4 where: 0 3 = l Xii (x i* * x *i> i=l N' r 04 = E X i j ( X j . * X . i ) 2 i= l j=l When N is large/ R is asymptotically normal confidence limits can be constructed addition/ (Cohen/ such that 1960). In this property can be utilized to test for significant differences between two independent KHATs by evaluating the normal curve deviate: Z " (*i - R 2 > V®12 + 522 If the test statistic/ Z/ exceeds the tabled value of the standardized normal distribution for the chosen significance level/ 1968). the di fference is considered significant (Kirk/ A FORTRAN computer program has been published to compute KHATs and test for significance between pairs of error matricies and 1982). (Congalton/ 1981/ and Congalton et a l ., 1981 Although this program has been used by other authors (e.g. Benson and DeGloria/ 1985)/ in the course of 47 this study it was found to contain an error and should not be utilized as published. written Therefore/ a PASCAL program was for the IBM-PC (Appendix D) which calculates a KHAT and its variance for a given error matrix. Chapter IV RESULTS AND DISCUSSION A. Classification Error Analysis 1. Introduction An important element of remote sensing cla ssifications has been the development of statistical sampling, determining accuracies/ (Draeger and Carneggie/ Scherk, 1975/ 1974/ procedures and hypothesis Hixson/ and Head and Szajgin, 1981/ for testing Kalensk y and 1982 and 1982). Since this study utilized complete enumeration of the test sites/ a ma p sampling scheme was unnecessary. A number of authors have addressed the problem of sampling and determining the confidence interval of the map acc uracy (Aronoff/ Fi t z p a t r i c k - L i n s / 1981, Rosenfield, Hay, et a l ., 1982, Van Genderen, 1979, 1981/ Hord and Brooner, Rosenfield and Melley, 1976, 1980, and et a l ., 1977). A variety of procedures and measures have been employed for reporting accuracies. informative, The easiest, and possibly least measure of acc uracy uses relative proportions. This technique simply compares the proportion of each class in the classified scene with a reference source, provides a measure of accuracy (Hixson, al., 1978, and Mazade, 1981). without regard to but neglects compensating 48 Hoffer, This procedure may be utilized to obtain area estimates, positional accuracy, 1981, which then et 49 classification errors and their effect on classification accuracy (Hixson, 1981). A wid el y accepted method for presenting the results of a classification is the confusion table or e r r o r - m a t r i x . These tables are a two-way cross-tabulation comparing the Landsat classification with reference data aerial photo interpretation, from maps, or ground verification. differing slightly in design, Although Table 2 illustrates the format and definition of the major components of a confusion table. In an effort to provide a single estimate of the classification accuracy of individual classes, approaches have been Hoeffer, 1979, Williams, followed. Some authors Roller and Visser, 1976) 1980, two (Fleming and Walsh, 1980, and have utilized the ratio of the number of correct points for a particular class to the total count of that class as determined by the Landsat classification, measure of individual class accuracy. and Shasby, 1981, 1984, Shasby, Hixson, 1981, et a l ., 1981, Other authors Kalensky, and U.S. et al., as a (Burgan 1979 and Forest Service, 1983) have utilized the ratio of the number of correct points for a particular class to the total count of that class as determined from the reference data. Since these single estimates ignore omission and commission errors resp ec ­ tively, a number of authors (Fox and Mayer, a l . , 1983, Kalensky, Fox, 1981, Mayer, 1979, Fox, et 1974, K r u m p e , et a l ., 1973, Mayer and et a l ., 1979, Mead and Meyer, 1977, 50 Table 2. Confusion table format. LANDSAT CLASSIFICATION COVER TYPE (MAP) Redl Pine Red Pine^ Jack Pine Swamp Percent® Pine Mixtures Conifers Other® Total® Correct 4 Jack Pine Pine Mixtures Swamp Conifers Other Tot a l 6 [ I7 Percent Correct® r 1 - columns correspond to classes/ as determined Landsat classification, l 10 from the and show which cover type classes they actually represent 2 - rows correspond type map, to classes, as determined from the cover and show into which Landsat classes it was placed 3 - includes hardwoods and all non-forest categories 4 - values along the diagonal represent correctly classified pixels 5 - the total count for a class from the cover type map 6 - the total number of pixels for a particular Landsat class 7 - the total number of pixels for the entire sample Table 2 (c o n t 1d .). 8 - the accuracy for a single class, errors only, ratio tions for that row considering omission of the number of correct classifica­ to (expressed as a the row total percent) 9 - the accuracy for a errors only, ratio single class, considering commission of the number of correct cl assifi ca­ tions for that column to the column total (expressed as a percent) 1 0 - overall classification accuracy, ratio of the sum of diagonal values to the total number of sample points (expressed as a percent) 52 Salazar/ 1982/ and Werth/ 1981) report two estimates of accuracy/ one with respect to omission errors and the other with respect to commission errors. A commonly used measure of overall classification accuracy is the ratio of diagonal values (i.e. correct classifications) sample points et al., (Bryant, 1978, Werth, et al., 1980, to the total number of Hixson, 1981/ and Williams, 1981, 1976). Strahler, All three measures of classification accuracy are reported in the confusion tables used in this study and will be discussed. 2. Unsupervised Clustering Output from the clustering wherein each pixel technique consists of a file is assigned to a cluster number. summary of the cluster statistics produced. (Table 3) A is also These statistics indicate the percentage of the scene classified into each cluster class and the mean brightness value, for each band, which characterize individual clusters. For example, the the data in Table 3 indicate that over 11 percent of the scene was classified into cluster number 26. prepared (Figure 10) From these data, showing band for each cluster. Next, the mean brightness value (pixel) file and the Of the 27 clusters created, assigned to the "other" cat egory and only five numbers 20, 21, 24, per a two-way cross-tabulation (Table 4) was run using the cluster cover type map file. a plot can be 25, and 26) 22 were (cluster to forest categories. of the forest clusters consisted of a mixture of forest Each 53 Table 3. cluster % 4 5 6 7 Clustering statistics, default parameters, Wexford County test site. #1 #2 3.510 69.79 82.76 81.31 63.63 # 6 # 7 2.557 86.38 103.15 98.76 74.68 4.768 47.59 54.12 55.92 45.93 4.231 65.15 76.90 75.91 60.17 7 #11 % 4 5 6 7 1.779 102.21 120.87 115.59 85.47 #16 % 4 5 6 7 6.668 117.80 126.95 126.68 93.30 #21 % 4 5 6 7 % 4 5 #10 3.089 80.03 95.66 92.13 70.63 1.469 94.63 112.75 106.99 79.71 1.769 90.78 108.12 103.70 77.92 #12 #13 #14 #15 3.107 111.76 126.69 124.07 90. 28 3.438 38.80 42.48 44.02 37.30 1.669 98.75 117.21 110.92 82.08 5.366 55.35 64.75 65.31 52.97 #17 #18 #19 3.255 74.72 89.05 86.33 66.80 2.216 107.02 125.29 118.73 86.87 4.555 43.68 49.00 52.79 44.51 #23 #24 #25 3.271 35.13 37.62 40.20 34.61 4.935 24.37 23.35 34.40 33.87 2.634 32.47 33.84 44. 20 40.70 6 7 #22 2.881 37.74 40.64 48.87 43.19 3.458 42.20 46.81 48.53 40.35 #26 #27 11.202 18.31 16.01 29.62 31.99 5.073 60.04 70.67 70.48 56. 56 1.677 27.50 27.83 31.00 25.83 I 6 5.589 51.25 59.08 60.52 49.40 #5 # 9 1 % 4 5 #4 CO 2.811 31.03 32.19 35.67 31.37 #3 #20 3.027 27.59 27. 55 39.78 38.40 54 N 0M « P o r * » i « i t n t 10 100 90 80 10 70 IT 50 40 30 20' 10 4 5 Spectral Figure 10. 6 7 Bend Plot of mean brightness values per band for 27 clusters. 55 Table 4. Two-way cross tabulation, default clusters, Wexford County test site Cover Type Hap Cluster Number 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 Red Pine Jack Pine 27 85 81 189 127 37 114 51 16 23 146 18 40 46 30 6 3 0 1 8 50 14 144 45 2 0 21 0 46 3 197 608 346 5 23 13 0 66 42 31 102 27 68 64 4 21 6 6 17 76 27 103 5 20 902 389 1921 3 11 61 16 3 4 1 52 93 121 49 582 123 Pine Mixtures 12 10 20 23 16 8 20 13 3 7 3 0 7 3 0 209 65 533 7 Swamp Conifers 26 0 1 2 0 0 3 0 0 0 0 0 9 0 3 0 0 0 12 Other 318 567 742 926 882 480 735 541 334 370 462 862 465 403 907 1882 529 590 659 90 239 536 373 143 158 108 158 Cluster Assignmen other other other other other other other other other other other other other other other other other other other red pine red pine other other red pine red pine red pine other 56 types with red pine the largest single type in each cluster. Cluster number 26 was the purest forest class (96.7 percent)/ whereas cluster number 20 most nearly represented a single forest type (70.9 percent red pine). from Table 4, a Aggregating the cluster assignments confusion table^ was prepared pine pixels (Table 5). Of the 5677 red in the scene, 4183 were correctly classified/ 894 were classified as other/ while 2801 pixels classified as red pine were/ in fact/ other forest types coniferous forest (263). (CE) only. red pine or non- Red pine was thus classified with an accuracy of 84.3 percent/ (OE) only/ (2538) considering omission errors or 63.1 percent considering commission errors The other (2538 pixels/ 25.2 percent). forest types were either classified as 74.8 percent) The non-coniferous was somewhat more accurate (98.1 or other (856 pixels/ forest cat egory (other) (OE) or 88.7 percent (CE)) such that overall classification accuracy was 88.7 percent. When coniferous forest is considered as a single category/ overall classification accuracy is 88.7 percent. To evaluate the spatial distribution of errors/ map was produced showing correctly classified (forest) pixels and pixels which were omitted and committed 11). Large blocks of omitted forest (e.g./ were gen erally stands with lighter stocking a (Figure A in Figure 11) levels/ from 25 All confusion table values were adjusted from photo interpretation of aerial photography as outlined in section IV.B. 57 to 50 percent crown closure (Figure 12). large blocks of committed forest, 1 2 , were In comparison the such as B in Figure 11 and frequently lowland sites dominated by brush, lowland hardwood species, white-cedar. and a sparse stocking of northern A large number of errors, both omission and commission, can be attributed to boundary positioning or the effects of numerous edges. Note especially the distribution of errors in the southwest portion of the site (Figure 11) where stands are smaller and irregularly shaped. Unsupervised clustering over the Crawford County test site, using default parameters, 6. is summarized in Table The cross-tabulation of clusters and cover type map categories assigned 17 clusters to the "other" category and ten clusters to forest types (Table 7). Cluster number 14 was the only forest cluster which was reasonably "pure," 96.8 percent forest, 79.3 percent swamp conifers. Each of the clusters assigned to jack pine consisted of a mixture of types; they ranged from 52.7 to 92.0 percent forest, only 39.7 to 65.6 percent jack pine. but While jack pine was classified with 88.0 percent accuracy (Table 8 ) considering omission errors only, the variable nature of the clusters included a large number of other forest types and non-forest pixels such that the accuracy, errors, was only 67.2 percent. cluster was reasonably "pure" percent considering commission Although the swamp conifer for an accuracy of 79.8 (CE), a large number of swamp conifer pixels were 58 Table 5. Landsat classification performance, clusters, Wexford County test site. default LANDSAT CLASSIFICATION COVER TYPE (MAP) Red Pine Jack Pine Pine Swamp Mixtures Conifers Other Total Percent Correct Red Pine 4183 894 5677 84.3 Jack Pine 1137 617 1754 — Pine Mixtures 1106 177 1283 — Swamp Conifers 295 62 357 — Other 263 13721 13984 Total 7584 15471 23055 Percent Correct 63.1 88.7 98.1 80. 3 59 f‘i r ’r': St.. . li «; . •’ $%/«V ; «* /•/ f/ t* w j:;1' & sir’ B. :H ni ut. SF* warnf •V:'-.;:-:i /i/f /;/e1* /rt/frt m /•rtf*. ...... rtiiffltti. '• i*; *!*‘: *» '*tS H i ■«:«: #*/£«: liiS am aH t i f . *•* '••rtf.!..}; 111 S tt* i :/'**:/.* ** .?•***!* *••• ■/•«*rt-rt::;;///:..*••;..,rt.;.. 'M * 7 ! t» •• J* i*>iiitMi,*>fi> *•»••**Ji■**** ILMililiilimi $ M ■!! & w ■ 1. 5*1 «*•••«••*»*•••* * •«* *«*»*«••*»*•••••• IM S 4 «**•«• «»•* • **•*••••••••••«** ■•»••«••*4**44 * **• X•*« * *•* >*»•*•*•*• *• «•**«l»* • » • • * » • « • • * • * * « * » • * • * ****** ,.•***#**•*»* «••*•**»*•*»**•** _«<-*••••• *»« •« *« *• + • *«***•«***• « •••****•» *•»*•* •• •• ••*»■*' ,••• **»» *» #« *•» •*•**•***••** *••ten** •• • ** *« ** **•«*•••••»*« •• «* »• •* '* «+««>** * ••*••*##**••#•***••»• ••• »•* • *•*••**■ >»«* ft* * * * * •« • * • •• * * * » * • * * * •* »«•« * * * *,***•»••**•*■" 4 »4 •V *4 4 4 Jiiimiii .4 ft* • • »• *•*• *» * ft* \ * • • # • • * « • • • « • • • * * • • « ! ■ •* \*«»«*«*« « * • « » •• * * * * •* » * "" »*•••**«•••••*«••** *•««»**• \*••••••••••*«»«•». iiHliliiiiiiif::: ** *W mm m m • •ww::; *-*- • * * * !* J *4 * I I I I • ■ ■ ■ II i « i i i i i i a« ! • • ■ II II ...... Figure 13. § FOREST CORRECTLY IDENTIFIED ***1 * * * * (tv h p a * 4* H I H i l l I* i | l l t i l l I* ■ * ■ ■ ■ r * k 4rft £**«**•**.** COMMISSION ERRORS OMISSION ERRORS Forest error map/ default clusters/ County test site. Crawford 66 Figure 14. Crawford County test site (portion), A-omission of lightly stocked jack pine. Figure 15. Crawford County test site (portion), B-commission of hardwood stand. 67 Tabl e 9. iter (•scene) % 4 Band 5 6 7 % 4 5 6 7 % 4 5 6 7 % 4 5 6 7 Cl ustering statistics/ smaller c luster radi Wexford County test si te. # 1 # 2 # 3 # 4 # 5 1.019 87.81 96.11 96.31 72.83 5.307 46.17 52.78 56.60 48.02 3.484 70.37 84.68 82.90 64.95 1.374 101.26 8 8 . 28 # 6 # 7 # 8 # 9 #10 6.706 42.79 47.78 49.68 41.43 3.597 36.09 38.22 48.54 44.82 .308 77.50 82.20 87.68 68.36 2.444 92.56 110.48 106.71 78.53 6.585 56.45 59.90 61.19 48.64 #13 #14 #15 2.531 74.31 88.57 85.09 6 6 . 24 4.621 35.38 37.96 40. 58 34.98 #11 #12 4.565 49.98 57.52 58.13 47.27 3.707 106.24 116.51 115.91 85.68 2.587 28.57 29.55 33.44 30.26 #16 #17 #18 3.127 80.18 95.63 92.16 70.74 2.064 122.43 127.00 126.95 97.21 1.176 15.82 12.14 21.08 2 2 . 10 #21 % 4 5 6 7 5.427 24.68 23.35 36.17 36.00 #26 % 4 5 6 7 2.691 31.00 30.00 35.00 31.00 #22 8.354 114.73 126.97 126.33 92. 27 #27 4.054 37.79 41.01 43.08 37.58 #23 3.615 65.05 77.00 74.49 58.49 110.88 110.31 81.68 #19 6.801 61. 59 72.91 72.88 58.13 #24 3.479 15.76 12.79 24.50 26.58 2.442 106.14 99.56 74.85 #20 4.336 19.18 17.68 32.59 34.99 #25 3.648 15.90 12.78 29.14 32.76 68 type map categories assigned 20 clusters to the "other" category and 7 clusters to forest types (Table 10). The "other" cat egory had few forest types associated with itr 15 clusters were 80 percent or more pure while 7 clusters were 90 percent or more pure, percent (OE) (Table 11). for an accuracy of 98.3 Commission errors within the "other" category were similar to the default cluster for an accuracy of 87.7 percent (OE). There were slightly fewer omitted and committed red pine pixels 80.4 (OE) or 70.3 percent number of pixels (572) (CE). Jack pine had a large omitted to the "other" category for an accuracy of only 37.9 percent (CE). for an accuracy of (OE), and 71.2 percent Overall classification accuracy was 82.2 percent whereas coniferous forest, as a single category, was classified with 90.6 percent accuracy. Comparison of the error map from the cluster with a smaller radius (Figure 16) to the map with the default cluster slight decrease (Figure 11) shows a in the omission and commission error pixels. Clustering with a smaller allowable cluster radius, the Crawford County test site, is summarized for in Table 12. The cross tabulation of clusters and cover type map categories assigned 17 clusters to the "other" cat egory and ten clusters to forest types (Table 13). Cluster 69 Table 10. Two-way cross tabulation/ smaller cluster radius, Wexford County test site. Cover Type Map Cluster Number Red Pine 1 2 25 288 3 4 5 86 6 7 15 28 84 606 Jack Pine 6 6 0 53 9 10 2 6 0 0 0 24 82 7 15 18 1 5 108 46 21 2 9 27 165 77 9 7 51 42 22 19 145 48 182 17 136 18 2 15 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 201 877 1182 4 44 350 916 85 26 Swamp Conifers 69 18 8 10 11 12 Pine Mixtures 8 21 0 0 3 9 4 29 7 54 34 0 10 0 0 0 277 47 91 49 27 29 165 205 7 51 50 1 0 12 0 1 304 91 31 15 58 4 29 18 0 36 259 7 126 79 2 0 0 39 0 Other 178 719 556 324 497 1067 181 41 531 1143 763 973 246 422 494 541 532 30 1147 69 94 2384 649 51 4 287 536 Cluster Assignment other other other other other other red pine other other other other other jack pine other other other other jack pine other red pine red pine other other red pine red pine other other 70 Table 11. Landsat classification performance, smaller cluster radius, Wexford County test site. LANDSAT CLASSIFICATION COVER TYPE (MAP) Red Pine Jack Pine Pine Swamp Mixtures Conifers Other Total Percent Correct Red Pine 4472 50 1040 5562 80.4 Jack Pine 574 699 522 1845 37.9 Pine Mi xtures 961 76 229 1266 — Swamp Conifers 214 59 86 359 — Other 142 98 13783 14023 Total 6363 982 15710 23055 Percent Correct 70.3 71. 2 — — 87.7 98. 3 82. 2 71 iiiisii ..iff/**/.-' /«/:•* ‘II! ***« * Si •"* . , r s i £ •* ■it; «L W * M m s * Wiim :« » m * •»» .1 - - m 'ir- 1 1 I W S -, ** .. ■**.» ■ , //ifJjm WiSgfr .-! » * ***• if, ***» *•* !***••••:* !ii&" •f • I* . ® S ii Vit'HtS:::!! it:;!:: ptiiiHi tttiiiinttt liiiiiilllii! FOREST CORRECTLY COMM ISSION ER R O R S IDENTIFIED tin m i OMISSION ER R O RS ...... Figure 16. Forest error map, smaller cluster radius, Wexford County test site. 72 Table 12. cluster % 4 5 6 7 % 4 5 6 7 #1 #2 #3 #4 #5 6.781 54.65 61.48 64.30 51.84 .219 103.91 111.65 112.40 86.55 .193 98.11 105.00 106.11 82.50 .722 91.26 102.80 102.17 79.33 3.038 90.98 107.96 105.66 79.61 # 6 # 7 # 8 # 9 #10 5.968 44.93 49.89 53.74 44.26 2.756 105.87 125.79 119.94 89.92 5.210 75.73 90.55 88.73 6 8 . 30 4.047 85.16 101.60 98.95 75.43 .522 113.04 121.31 121.57 92.66 #13 #14 #15 5.027 30.86 32. 20 38. 27 34.00 #11 % 4 5 6 7 % 4 5 6 7 6 7 6.832 64.72 75.02 74.93 58.95 5.622 72.37 83.43 64.72 .278 107.73 119.48 119.64 92.89 #16 #17 #18 #19 6.520 48.77 55.38 58.76 47.80 2.061 24.83 24.36 33. 38 30.98 2.374 115.36 126.95 126.84 97.93 6.503 37.72 41.06 46.00 39.46 #26 % 4 5 6 7 #12 2.775 99.08 119.55 114.47 84.67 #21 % 4 5 Clustering statistics/ smaller cluster radius, Crawford County test site. 2.503 119.00 127.00 127.00 96.33 3. 140 28.36 28.50 34. 32 30.99 #22 3.429 22.29 21. 29 30.12 29.88 #27 .605 26.00 30.00 35.00 30.00 86.10 #20 7.195 60.40 70.74 70.97 57.00 #23 #24 #25 5.503 34.35 36.79 42.50 36.97 5.983 42.37 46.75 51.08 42.08 4.195 51.70 60.44 61.49 48.95 73 Table 13. Two-way cross tabulation, smaller cluster radius, Crawford County test site. Cover Type Map Red Pine Jack Pine 290 38 15 3 4 5 11 0 2 1 1 1 1 1 1 1 0 0 0 1 6 23 7 0 8 3 9 1 0 0 12 6 0 347 4 92 44 Cluster Number 1 2 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 4 10 Pine M ixtures 3 39 3 21 19 0 0 9 223 146 5 23 Swamp Conifers 42 0 0 2 0 0 7 0 22 0 1 0 53 789 109 118 0 0 0 0 18 27 39 14 26 544 318 292 264 821 281 65 38 43 28 64 35 116 35 146 9 740 475 145 109 40 14 21 44 23 8 0 2 0 0 125 9 86 674 90 52 12 0 10 Other 428 13 20 54 291 331 250 507 418 37 211 523 467 25 138 197 80 355 33 441 298 61 206 335 235 183 13 Cluster Assignment other other other other other jack pine other other other other other other other other jack pine other jack pine jack pine jack pine other jack pine swamp conifers jack pine jack pine other other jack pine 74 assignments were very similar to those from the default cluster {Table 7)/ with the swamp conifer class again the only nearly "pure" forest cluster. Accuracies obtained with the smaller cluster radius {Table 14) are likewise to those from the default cluster (Table 8 ). similar Omission errors increased slightly for the forest types, whereas commission errors increased for the "other" category. Overall classification accurac y was 73.5 percent and coniferous forest, accurate. The error map for the smaller cluster radius (Figure 17) compared as a single category, was 86.9 percent indicates only very minor differences when to the default cluster error map {Figure 13). In an effort to minimize the effect of a large number of non-forest snow covered pixels and subsequent clusters, data. level "other" slicing was used to preprocess the Landsat An analysis of previous cluster assignments indicated a maximum brightness value, per band, pixels were unlikely to occur. above which forest All pixels above this threshold were therefore assigned a value of zero and excluded from further consideration. unsupervised clustering, sliced scene, Table 15. The results of the using default parameters on a level for the Wexford County test site are shown The 27 clusters thus produced are more tightly confined due to the smaller brightness range displayed by the level sliced scene. The cross tabulation of clusters and cover type map categories assigned 17 clusters to the in 75 Table 14. Landsat classification performance/ smaller cluster radius/ Crawford County test site. LANDSAT CLASSIFICATION COVER TYPE (MAP) Red Pine Jack Pine Pine Swamp Mixtures Conifers Other Total Percent Correct Red Pine — 323 — 24 38 385 Jack Pine — 5746 — 319 784 6849 — 716 — 40 114 870 Swamp Coni fers — 830 — 707 30 1567 45. 1 Other — 637 — 22 4275 4934 86.6 Total — 8252 — 1112 5241 14605 Percent Correct — — 63.6 81.6 Pine Mixtures 69.6 — 83.9 — 73.5 76 Jimimmwmu'm I#:#::;:::: ::•• * .\\*>■.':■ • f.® m m i s i H m m . n il® j m k W M :/*vv ;v/::7A? i m .smii i'iiiilim ■ M i i i i m m i i i i : m w m r m limm im W i :w m M a M M t m e i a M m m I 'HisiisS: m M M •••m m TTI FOREST CORRECTLY !!!! ! !>*<•< IDENTIFIED M It •• •• i 4 ■» • • «« t i » 2 V M m i iM iin c o m m is s io n errors OMISSION ERRORS Figure 17. Forest error map/ smaller cluster r a d i u s ; Crawford County test site. 77 Table 15. cluster % 4 5 6 7 % 4 5. 6 7 #1 6 7 % 4 5 6 7 7.032 46.11 52.18 55.78 46.78 # 6 # 7 .847 0.00 0.00 79.63 61.86 7.485 55.00 64.44 65.13 53.16 6 7 6 7 #12 8.484 15.87 13.00 27.88 31.51 1.853 45.86 47.00 44.86 31.86 # 8 .432 68.31 #4 .772 0.00 0 . 00 0.00 62.23 # 9 #5 4.191 32.71 84.70 36. 38 31.86 #10 0.00 2.03 15.23 11.59 19.18 62.34 21.01 #13 #14 #15 1.704 30.03 31.20 43.48 41.15 1.498 66.39 78. 11 79.80 2. 285 42.52 47.44 54.00 45.75 0.00 0.00 0 .00 .978 67.83 80.23 0.00 #17 #18 #19 4.559 21.98 20.49 28.34 27.39 1.984 38.74 41.77 53.86 47.30 6.660 21.42 19.60 33.96 35.61 6.412 60.95 72.95 71.00 56.64 3.185 65.02 76.97 74.48 58.48 #23 #24 #25 3.447 33.40 34.96 47.04 43. 21 7.166 51.29 59.21 62.57 51.80 8.782 42.37 47.73 50.25 41.71 2.150 46.26 51. 35 61.12 52.68 #26 % 4 5 1.768 66.97 79.58 78.76 61.36 #3 #16 #21 % 4 5 #2 5.055 31.95 33.68 41.28 37.77 #11 % 4 5 Clustering statistics, default parameters on level sliced scene, Wexford County test site. 4.361 27.43 27.03 38.66 37.69 #22 1.187 42.49 46.05 57.67 50.69 #27 3.688 31.00 30.00 35.00 31.00 #20 78 "other" category and 10 clusters to forest types 16). By reducing scene, more (Table the number of non-forest values in the forest clusters were created, 8 red pine and 2 jack pine, than when the entire brightness range was analyzed. The classification accuracy of red pine, (OE) or 74.2 percent (CE), is similar to previous clusters whereas jack pine was identifiable with 50.3 percent (CE) 81.4 The "other" category was classified with 97.7 percent accuracy, considering omission errors only, only. accuracy (Table 17). (OE) or 58.7 or 89.0 percent considering commission errors Overall classification accuracy was 83.0 percent whereas coniferous forest, considered as a single category, was classified with 91.4 percent accuracy. forest clusters were created, Although more accuracies are within 3 percent of those obtained with the previous classifications. Unsupervised clustering results using default parameters on the level sliced scene for the Crawford County test site are shown in Table 18, The cross tabulation of clusters and cover type map categories assigned 15 clusters to the "other" category and 12 clusters to forest types 19). (Table The swamp conifer clusters were reasonable "pure," 80.7 to 94.3 percent conifers, forest and 61.5 to 80.2 percent swamp such that the accuracy was 74.5 percent, considering commission errors only (Table 20). Due to a large number of pixels being classified as jack pine (976) , accuracy considering omission errors was only 37.3 percent. 79 Two-way cross t a b u l a t i o n , default parameters on level sliced scene, Wexford County test site. Table 16. Cover Type Map Cluster Number 1 2 3 4 5 Red Pine 106 37 7 1 Jack Pine Pine Mixtures 154 62 5 4 52 56 3 5 1 0 0 0 0 0 10 0 0 7 19 7 9 9 4 8 2 1 9 42 33 113 1344 271 87 70 135 226 1140 82 27 165 113 299 145 17 635 38 4 313 40 28 5 2 6 6 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 Swamp Conifers 6 40 384 13 24 31 24 14 4 59 47 108 33 156 69 16 327 19 14 25 128 33 168 12 4 23 14 51 47 3 1 0 86 25 939 2 0 24 139 258 195 98 42 831 443 181 57 173 887 980 84 272 5 75 3 27 0 1 1 3 23 2 13 17 38 33 25 399 845 190 105 388 98 209 50 18 3 31 20 120 Other 12 Cluster Assignment other other other other other other other other other jack pine other red pine red pine other other jack pine red pine red pine other other red pine red pine red pine other other red pine other 80 Table 17. Landsat classification performance/ default parameters on level sliced scene/ Wexford County test site. LANDSAT CLASSIFICATION COVER TYPE (MAP) Red Pine Jack Pine Pine Swamp Mixtures Conifers Other Total Percent Correct Red Pine 4654 175 888 5717 81.4 Jack Pine 339 901 551 1791 50. 3 Pine Mixtures 857 267 174 1298 — Swamp Conifers 187 114 63 362 — Other 239 77 13571 13887 Total 6276 1532 15245 23055 Percent Correct 74.2 58.7 89.0 97.7 83.0 81 Table 18. Clustering statistics, default parameters on level sliced scene, Crawford County test site. cluster % 4 5 6 7 % 4 5 6 7 #1 6 7 6 7 # 6 # 7 # 8 # 9 7.705 49.32 56.37 59.25 48.16 7.709 42.14 46.57 50.94 42.38 7.175 35.10 37.44 42.68 36.77 7.756 38.63 42.05 46.81 39.56 3.420 24.44 23.62 30.48 28.51 2.835 75.46 89.93 88.47 68.26 #21 % 4 5 6 7 1.497 20.84 19.28 26.87 26.82 #26 % 4 5 6 7 #4 5.290 55.53 64.68 66.14 53.07 #16 % 4 5 #3 1.646 71.28 84.49 87.35 67.54 #11 % 4 5 #2 .422 0.00 91.14 88.90 68.54 #12 1.145 14.92 12.45 21.36 25.31 #17 3.977 61.46 72.34 72.16 57.12 #22 .720 0.00 0.00 90.05 69.46 #27 .828 76.00 91.00 0.00 0.00 3.816 58.51 68.69 68.58 54.56 3.101 66.90 79.07 77.88 61.08 #5 3.562 63.58 74.70 75.56 59.59 #10 5.770 27.71 27.99 34.57 31.22 #13 #14 #15 6.726 31.38 32.61 3B.85 34.11 8.178 45.75 51.37 55.18 45.26 3.082 68.95 81. 57 82.01 64.10 #18 2. 273 72.60 86.16 83.43 65.12 #23 .837 18.16 15.93 23.64 24.04 #19 #20 89.93 68.79 2.156 58.15 67.55 72.33 57.65 #24 #25 .415 76.92 0.00 6.490 52.67 61.00 62.87 50.74 1.471 64.28 75.76 80.12 63. 26 82 Table 19. Two-way cross tabulation, default parameters on level sliced scene, Crawford County test site. Cover Type Map Cluster Number 1 2 3 4 5 Red Pine 3 3 4 54 165 1 52 107 270 465 784 752 767 400 3 5 6 21 7 25 37 25 49 32 8 9 10 11 12 0 13 14 15 16 17 18 19 51 29 20 21 5 22 23 24 25 26 27 Jack Pine 86 Pine Mixtures 8 27 11 2 12 37 47 105 58 86 54 3 Swamp Conifers Other 3 4 34 56 85 82 172 234 195 116 280 228 215 209 330 351 233 297 119 62 42 126 38 149 381 211 0 11 2 121 0 1 0 2 0 0 8 848 386 70 36 98 32 3 96 93 9 14 5 216 0 1 9 13 0 2 135 51 36 14 5 235 53 20 0 0 2 1 1 12 0 0 0 295 89 26 69 2 2 7 1 0 16 34 8 12 9 7 9 0 231 209 138 36 80 20 Cluster Assignment other other other other other jack pine jack pine jack pine jack pine jack pine jack pine swamp conifers jack pine jack pine other other other other other jack pine swamp coni fers other swamp coni fers others other other other 83 Table 20. Landsat classification performance/ default parameters on level sliced scene/ Crawford County test site. LANDSAT CLA SSIFICATION COVER TYPE (HAP) Red Pine Jack Pine Red Pine Jack Pine Pine Swamp Mixtures Conifers Other Total — 339 11 33 383 — 7728 127 692 8547 102 830 Percent Correct — 90.4 Pine Mixtures — 705 23 Swamp Conifers — 976 595 24 1595 37.3 Other — 887 43 2320 3250 71.3 Total — 10635 779 3171 14605 Percent Correct — 72.7 74.5 73.2 — 72.9 84 Jack pine suffered from the opposite problem: a large number of commission errors produced an acc uracy of 72.7 percents and a lesser number of omission errors produced an ac curacy of 90.4 percent. The "other" category was classified with an accuracy of 71.3 (CE) (OE) or 73.2 percent for an overall classification ac curacy of 72.9 percent. Due to the large number of forest pixels being classified as "other/" a single category/ the ac curacy of coniferous forest/ was only 87.8 percent. increasing the accuracy of jack pine/ as Although this clustering procedure resulted in an overall loss of accuracy compared to previous clustering results. To complete the series of unsupervised clusters/ the clustering routine with smaller allowable cluster radius (from seven to three digital counts) was run on the level sliced scene of the Wexford County test site (Table 21). The cross tabulation of clusters and cover type map categories assigned 17 clusters to the "other" category and 10 clusters to forest types (Table 22). The 8 red pine clusters were reasonable "pure/" 57.9 to 99.3 percent and 49.4 to 81.4 percent red pine forest for an accuracy of 82.7 percent considering omission errors only (Table 23). Commissions/ e spe cially including pine mixtures and other/ lowered the accurac y to 76.0 percent represented by 2 clusters, errors of omission, (OE). Jack pine/ was quite evenly balanced between esp ec ia lly to the other category, and 85 Table 21. Clustering statistics, smaller cluster radius on level sliced scene, Wexford County test site. cluster % 4 5 6 7 % 4 5 6 7 #1 6 7 % 4 5 6 7 6 7 % 4 5 6 7 #4 7.032 46.11 52.18 55.78 46.78 1.853 48.86 47.00 44.85 31.86 62.28 # 6 # 7 # 8 # 9 .847 0.00 0.00 79.63 61.86 7.483 55.00 64.44 65.13 59.16 1.768 66.97 79.58 78.76 61.36 #12 8.484 15.87 18.00 27.88 31.51 .432 88.31 0.00 0.00 0.00 .772 0.00 0.00 0.00 .878 67.83 80.23 0.00 #5 4.191 32.71 84.70 36.88 31.86 #10 2.037 15. 23 11.59 19.18 62.34 21.01 #13 #14 #15 1.704 80.03 31. 29 48.48 41.15 1.498 66.39 78.11 79.80 2. 285 41.52 47.44 54.00 45.75 0.00 #16 #17 #18 #19 4.559 21.98 20.49 28.34 27.39 1.984 38.74 41.77 53.86 47.30 6.660 21.42 19.60 33.96 35.61 6.412 60.95 72.95 71.00 56.64 8.185 65.02 76.97 74.48 58.48 #23 #24 #25 3.447 33.40 84.96 47.04 43.21 7.166 51.29 59.21 62.57 51. 80 8.782 42.37 47.73 50.25 41.71 #21 % 4 5 #3 5.055 31.95 33.68 41.28 37.77 #11 % 4 5 #2 #22 2.150 46.26 51.35 61.12 52.68 1.187 42.49 46.05 57.67 50.69 #26 #27 4.361 27.43 27.03 38.66 37.69 31.00 30.00 35.00 31.00 8.688 #20 86 Table 22. Two-way ctoss tabulation, smaller cluster radius on level sliced scene, Wexford County test site. Cover Type Hap Cluster Number 1 2 3 4 5 Red Pine 106 37 7 1 7 19 7 9 8 9 6 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 Jack Pine Pine Mixtures Swamp Coni fers 154 62 5 4 52 56 3 5 0 0 0 0 10 0 0 1 0 9 4 2 1 42 33 113 1344 271 87 70 135 226 1140 82 27 165 113 299 145 17 635 38 4 313 40 28 5 2 6 6 40 384 13 24 31 24 14 4 59 47 108 33 156 69 16 327 19 14 25 128 33 168 12 4 23 14 51 47 3 1 0 86 25 939 2 0 24 139 258 195 98 42 831 443 181 57 173 887 980 84 272 5 75 3 27 0 1 1 3 23 2 13 17 38 33 25 399 845 190 105 388 98 209 50 18 3 31 20 120 Other 12 Cluster Assignment other other other other other other other other other jack pine other red pine red pine other other jack pine red pine red pine other other red pine red pine red pine other other red pine other 87 Table 23. Landsat classification performance/ smaller cluster radius on level sliced scene/ Wexford County test site. LANDSAT CLASSIFICATION COVER TYPE (MAP) Red Pine Red Pine 4734 194 793 5721 82.7 Jack Pine 239 951 597 1787 53.2 Pine Mixtures 869 233 193 1295 — Swamp Conifers 156 117 87 360 — Other 235 78 13579 13892 Total 6233 1573 15249 23055 Percent Correct 76.0 Jack Pine 60.5 Pine Swamp Mixtures Conifers Other Total 89.0 Percent Correct 97.7 83.6 errors of commission percent for accuracies of 53.2 (CE) respectively. types such that the (OE) and 60.5 Most errors were among forest "other" cat egory had accuracies of 97.7 (OE) and 89.0 percent (CE). Overall classification accuracy was 83.6 percent while coniferous forests/ as a single category/ were classified with 91.4 percent accuracy. Results of clustering/ using a smaller cluster radius on the level sliced scene/ for the Crawford County test site are shown in Table 24. The cross tabulation of clusters and cover type map categories assigned 18 clusters to the "other" category and nine clusters to forest types 25). While cluster number nine was almost pure swamp conifers/ percent 95.8 percent/ the second cluster/ number 19/ 97.3 forest/ was less than half swamp conifers. Accuracies for the swamp conifer category were 51.0 (Table (OE) or 58.3 percent (CE) (Table 26). therefore All 7 jack pine clusters were composed of a mixture of categories/ pine mixtures, swamp conifers, and other, including such that classification accuracy was only 69.7 percent (CE). Most jack pine omissions were to the "other" category for an accuracy of 80.4 percent (OE). The "other" cat egory was well balanced between errors of omission and commission accuracies of 86.1 (OE) and 80.7 percent (CE). for Overall classification acc uracy was 72.8 percent while coniferous 89 Table 24. Clustering statistics, smaller cluster radius on level sliced scene, Crawford County test site cluster % 4 5 6 7 % 4 5 #1 #2 4.646 40.11 44.02 48.67 40.92 1.240 56.80 65.40 75.02 62.60 # 6 # 7 .207 77.19 .566 #3 .291 77.31 0 .00 0.00 0.00 # 8 .445 0.00 0.00 6 0.00 0.00 0.00 0.00 0.00 7 70.15 70.32 90.47 69.60 #11 #12 #13 % 4 5 6 7 % 4 5 6 7 2.618 76.41 91.45 89.05 68.89 92.33 #16 #17 #18 4.038 55.61 7.371 42.69 47.94 51.47 42.75 6.180 64.96 78.52 79.15 61.81 66.00 64.93 51.85 #21 % 4 5 6 7 % 4 5 6 7 . 298 0.00 0.00 0.00 #22 7.751 34.42 36.89 42.44 36.86 4.042 50.73 58.36 56.36 46.82 #26 #27 4.441 62.67 74.67 73.67 57.83 2.035 41.00 47.00 49.00 41.00 5.751 71.78 84.91 84.91 66.49 #23 6.392 29.47 30.82 36.47 33.41 #4 9.162 47.55 54.03 57.06 46.76 # 9 1.070 14.27 11.69 20.46 25.56 #14 4.788 27.46 27.83 33.90 30.48 #19 4.080 22. 50 21.70 30.40 29.73 #24 5.306 58.29 68.93 6 8 . 29 54.21 #5 .373 0.00 0.00 91.28 0.00 #10 .648 76.72 91.45 0.00 70.04 #15 3.387 61.80 73.25 71.52 56.73 #20 5.702 37.93 40.98 45.12 38.36 #25 7.168 52.40 63. 20 63. 20 51.80 90 Table 25. Two-way cross tabulation/ smaller cluster radius on level sliced scene, Crawford County test site. Cover Type Map Cluster Number Red Pine 1 2 23 3 3 4 5 36 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 0 0 0 0 0 1 0 2 0 6 36 6 2 21 6 34 15 53 4 53 7 16 6 0 Jack Pine Pine Mixtures 393 33 3 427 3 3 3 6 7 12 37 7 126 675 85 33 43 2 1 1 0 44 44 0 1 0 0 0 0 3 0 1 1 12 0 16 111 88 10 20 407 153 364 48 19 49 617 42 133 9 99 19 25 866 85 831 152 274 119 96 Swamp Conifers 12 10 319 0 0 0 1 Other 207 59 33 415 29 15 39 34 5 52 210 27 388 171 86 1 156 207 312 410 72 7 48 4 422 60 115 15 126 8 12 4 12 202 220 183 130 296 329 274 113 Cluster Assignment jack pine other other jack pine other other other other swamp conifers other other other other jack pine other other jack pine other swamp conifers jack pine jack pine other jack pine other other other other 91 Table 26. Landsat classification performance, smaller cluster radius on level sliced scene, Crawford County test site. LANDSAT CLASSIFICATION COVER TYPE (MAP) Red Pine Red Pine _ _ Jack Pine Pine Mixtures Jack Pine Pine Swamp Mixtures Conifers Other Total Percent Correct 294 39 40 373 — 5348 426 876 6650 — 629 59 123 811 Swamp Conifers — 707 781 43 1531 51.0 Other — 696 34 4510 5 240 8 6 .1 Total — 7674 1339 5592 14605 Percent Correct — 69.7 58.3 80.7 — 80.4 — 72.8 92 forest, as a single category, was classified with 87.6 percent accuracy. Overall classification accuracy was within one percent of both previous classifications. 3. Supervised Classification The final set of training site signatures for the Wexford County test site consisted of nine signatures red pine, five for jack pine, for "other" (Table 27). two for pine mixtures, After numerous attempts, for and six swamp conifers were dropped as a category since a "unique" signature could not be extracted from the scene. Classification of the pixels utilized to define training site signatures is a standard technique signature accuracies. for determining Training sites were thus classified utilizing the minimum dis ta nce -to-means and maximum likelihood classifiers (Table 28 and 29). accuracies are high, by design, Classification and should not be utilized as a measure of overall classification accuracy as reported in a few cases. With the exception of pine mixtures, the training sites were excellent representatives of the cover types and therefore were used to classify the entire test site. A cross tabulation of signatures and cover type map categories was developed from the minimum dis tance-to-means classifier, Signatures representing the same (Table 30). cover type were aggregated from Table 30 to produce a confusion table (Table 31). Both red pine and jack pine had 93 Table 27. Training site signatures/ Wexford County test site. Signature Cover Name Type PRl PR2 PR3 PR5 PR 6 PR7 PR 8 PR9 PR10 PJ1 PJ2 PJ3 PJ5 PJ 6 MIX3 MIX7 N2 N3 N4 N5 N8 N9 red pine red pine red pine red pine red pine red pine red pine red pine red pine jack pine jack pine jack pine jack pine jack pine pine mixtures pine mixtures other other other other other other Sample Size 24 24 35 30 30 16 36 60 20 36 18 28 20 18 35 41 30 32 71 34 40 20 Mean BV ---------------6 5 4 14.6 15.9 18.1 14.9 30.4 58.4 26.6 23.3 32.0 16.6 15.2 19.4 23.8 19.2 17.0 15.4 31.9 43.0 61.3 55.6 43.9 73.2 11.0 13.8 16.2 11.5 31.2 66.8 26.6 21.5 33.1 13.4 12.3 18.0 23.0 16.8 13. 5 12.6 32.7 46.9 73.1 65.8 49.3 87. 2 7 25.7 28.3 31.7 25.9 43. 2 79.3 38.9 36.4 44.6 30. 1 31.4 34.4 30. 1 40.9 65.6 37.2 36.6 41.5 21.0 21.6 21.2 19.7 26.8 31.1 25.8 24.7 23.7 34.1 46.2 72.9 65.8 51.8 83.6 25.9 29.0 25.2 27.6 26.3 30.4 36.5 57.8 52.6 44.0 65.4 94 Table 28. Training site classification, minimum distance to-means classifier, Wexford County test site. LANDSAT CLASSIFICATION Training Red Pine Site Jack Pine Pine Swamp Coni fers Other Total Mixtures Percent Correct 259 1 4 — 11 175 94.2 Jack Pine 2 114 1 — 3 120 95.0 Pine Mixtures 10 19 47 — 0 76 61.8 Swamp Conifers 0 0 0 — 0 0 Other 1 3 0 — 234 238 Total 272 137 52 — 248 709 95.2 83. 2 90.4 — Percent Correct — 98.3 < » l l vD 1 1• 1 ■ I Red Pine 92.2 95 Table 29. Training site classification/ maximum likelihood classifier/ Wexford County test site. LANDSAT CLASSIFICATION Red Training Pine Site Jack Pine Swamp Pine Mixtures Conifers Other Total Percent Correct Red Pine 268 3 4 — 0 275 97.5 Jack Pine 1 116 3 — 0 12 0 96.7 Pine M ixtures 9 7 60 — 0 76 78.9 Swamp Conifers 0 0 0 0 0 Other 0 1 0 — 237 238 Total 278 127 67 — 237 709 96.4 91.3 89.6 Percent Correct 100 — 9 9 ,6 96. 1 96 Table 30. Tw o-w ay cross tabulation/ minimum distance-tomeans classifier/ Wexford County test site. Cover Type Map Signature Name PR1 PR2 PR3 PR5 PR 6 PR7 PR 8 PR9 PR10 PJ1 PJ2 PJ3 PJ5 PJ6 MIX3 MIX7 N2 N3 N4 N5 N8 N9 Red Pine 287 533 741 188 280 116 499 738 468 2 7 58 132 9 60 42 22 23 95 294 494 279 Jack Pine 1 2 5 0 37 3 67 6 135 90 166 143 211 Pine Mixtures 22 0 83 75 23 42 13 113 79 85 4 13 60 131 3 5 127 20 11 114 97 31 156 94 64 89 165 77 Swamp Conifers Other 0 2 7 0 0 13 91 46 113 27 372 0 45 9 44 0 4 2872 10 22 17 6 8 21 21 46 97 59 5 24 15 1 0 12 10 49 210 19 6 0 525 812 1358 1903 1871 7026 97 Table 31. Landsat classification performance/ minimum di sta nce-to-means classifier/ Wexford County test si t e . LANDSAT CLASSIFICATION COVER TYPE (MAP) Red Pine Jack Pine Red Pine 4478 253 111 898 5740 78.0 Jack Pine 322 968 49 480 1819 53. 2 Pine Mixtures 616 274 226 176 1292 17.5 Swamp Conifers 136 137 41 49 363 Other 234 107 2 13498 13841 Total 5786 1739 429 15101 23055 Percent Correct 77.4 55.7 52.7 Pine Swamp Mixtures Conifers Other Total 89.4 Percent Correct — 97.5 83.1 98 similar omission errors, commission errors, individual including pine mixtures, accuracies were 78.0 for red pine and 53.2 pine. primarily to "other," and such that (OE) or 77.4 percent (OE) or 55.7 percent (CE) (CE) for jack The majority of pine mixture pixels were classified as red pine or jack pine; such that the pine mixture accuracy, considering omission errors, percent. The "other" category had accuracies of 97.5 or 89.4 percent 83.1 percent. was fairly low, 17.5 (OE) (CE) with overall classification accuracy of Coniferous forests, as a single category, were classified with an accuracy of 91.6 percent. A cross tabulation of signatures and cover categories, as obtained classifier, was prepared (Table 33). Although from the maximum likelihood (Table 32) as was a confusion table fewer pixels were committed to the red and jack pine classes, omitted, type map a larger number of pixels were particularly to the "other" category. were thus 75.4 (OE) or 79.1 percent 52.2 (OE) or 55.8 percent (CE) (CE) Accuracies for red pine and for jack pine. The pine mixture and "other" category had fewer omission and commission errors, classification. or 47.8 percent compared to the minimum distance-to-means Accuracies for pine mixtures were 22.5 (CE) the "other" category. and 97.9 (OE) or 8 8 . 8 (CE) (OE) percent for Overall classification accuracy was 83.1 percent whereas coniferous forest, as a single category, was classified with 91.3 percent accuracy. 99 Table 32. Two-wa y cross tabulation/ maximum likelihood classifier/ Wexford County test site. Cover Type Hap Signature Name PRl PR2 PR3 PR5 PR 6 PR7 PR 8 PR9 PR10 PJ1 PJ2 PJ3 PJ5 PJ6 MIX3 MIX7 N2 N3 N4 N5 N8 N9 Red Pine 354 501 Jack Pine Pine Mixtures Swamp Con i fers Other 1 2 28 69 0 2 0 1 688 3 7 98 412 74 582 712 280 4 3 26 181 66 8 5 0 64 3 67 116 123 94 254 8 122 10 74 72 30 3 17 23 75 1150 28 49 196 36 16 40 145 63 0 51 9 133 68 66 5 3 28 156 0 0 27 49 4 26 145 17 129 13 216 1 2 10 4 24 0 8 87 5 30 222 120 8 2 20 0 1 120 3 7 25 19 283 201 20 9 647 608 453 932 1255 9731 27 3 16 11 10 0 Table 33. Landsat classification performance, maximum likelihood classifier, Wexford County test site. LANDSAT CLASSIFICATION COVER TYPE (MAP) Red Pine Jack Pine Red Pine 4282 273 161 — 966 5682 75.4 Jack Pine 253 939 92 — 514 1798 52.2 Pine Mixtures 564 245 288 — 183 1280 22.5 Swamp Conifers 129 127 54 — 52 362 Other 187 98 7 — 13641 13933 Total 5414 1682 602 — 15356 23055 Percent Correct 79.1 55.8 47.8 Swamp Pine Mixtures Con ifers Other Total 88.8 Percent Correct — 97.9 83.1 101 The final set of training site signatures for the Crawford County test site consisted of nine signatures for jack pine, five for swamp conifers, (Table 34). and five for "other" Because of their small spatial extent, than 5 percent, less red pine and pine mixtures did not display di sti nguishable signatures and were therefore dropped as categories. Classification of the training site pixels by the minimum dis tance-to-means and maximum likelihood classifiers produced relativ ely high accuracies, 90.5 percent, but bel ow those obtained 88.3 and in Wexford County (Table 35 and 36). A cross tabulation of signatures and cover type map categories, as obtained classifier, was prepared co ntingenc y table from the minimum distance-to-means (Table 37) (Table 38). large number of commissions, conifers, and aggregated into a The jack pine class had a including pine mixtures, and "other," such that accuracy, commission errors only, was 68.9 percent. considering Jack pine had fewer omissions for an acc uracy of 86.2 percent. conifer type was reversed, percent accuracy, jack pine, swamp fewer commission errors and more omission errors, for an accuracy of 47.5 percent. The swamp for 68.7 particularly to Overall classification accuracy was 73.7 percent whereas the coniferous forest, accurate. as a single category, was 8 8 . 1 percent 102 Table 34. Training site signatures/ site. Crawford County test Mean BV Signature Cover Name Type PJl PJ2 PJ3 PJ4 PJ5 PJ6 PJ7 PJ8 PJ9 01 Q2 Q3 Q6 Q7 N1 N2 N3 N5 N6 jack pine jack pine jack pine jack pine jack pine jack pine jack pine jack pine jack pine swamp conifers swamp conifers swamp conifers swamp conifers swamp conifers other other other other other Sample Size 58 84 49 28 24 36 35 32 48 48 39 18 28 28 28 77 35 62 30 4 5 39.7 30.3 26.7 25.8 25.2 31.5 36.4 52.2 43.4 16.7 23.7 15.4 20.4 43.6 31.3 26.8 25.9 25.1 32.7 67.4 57.7 47.7 14.7 22.3 12.7 18.9 22.2 20.8 6 48.6 37.3 33.6 31.0 31.0 40.1 32.1 62.7 50.7 23.6 30.7 21. 1 25.9 29. 2 54.4 63.5 64.0 83.0 82.4 69.8 99.4 83.2 1 0 0 . 0 113.0 127.0 126.0 47.3 48. 5 42.0 7 40.5 32.6 29.9 28.5 27.9 35.3 52.6 50.5 41.6 26.7 29.9 25.4 27.0 29.1 51.9 64.4 75.7 96.6 39.7 103 Table 35. Training site c l a s s i f i c a t i o n r minimum distanceto-means classifier, Crawford County test site. LANDSAT CLASSIFICATION Training Red Site Pine Red Pine Jack Pine Pine Swamp Mixtures Conifers Other Total Percent Correct 0 0 0 0 357 25 16 398 0 0 0 0 — 32 137 0 169 81. 1 Other — 17 0 187 204 91.7 Total — 406 162 203 771 87.9 84.6 92.1 — Jack Pine — Pine Mixtures — Swamp Conifers Percent Correct — 89.7 — 88.3 104 Table 36. Training site classification, maximum likelihood classifier, Crawford County test site. LANDSAT CLASSIFICATION Red Training Pine Site Red Pine Jack Pine Swamp Pine Mixtures Conifers Other Total Percent Correct 0 0 0 0 362 22 14 398 0 0 0 0 — 23 146 0 169 86.4 Other — 14 0 190 204 93.1 Total — 399 168 204 771 90.7 86.9 — Jack Pine — Pine Mixtures — Swamp Conifers Percent Correct 93.1 — 91.0 — 90. 5 105 Table 37. Two -way cross tabulation; minimum distance-tomeans classifier/ Crawford County test site. Cover Type Map Signature Name Red Pine Jack Pine Pine Mixtures PJ1 PJ2 PJ3 PJ 4 PJ 5 PJ 6 PJ7 PJ8 PJ9 Ql Q2 Q3 06 07 N1 N2 N3 N5 N6 30 34 28 5 597 630 502 108 174 654 893 475 617 3 39 62 6 45 39 37 37 1 16 0 68 2 3 74 8 20 21 68 5 512 431 95 0 0 51 11 68 14 15 104 105 58 63 0 15 0 8 14 59 55 37 7 7 Swamp Conifers Othe 66 271 90 131 43 28 97 96 44 70 128 117 175 185 103 24 9 3 100 0 8 64 12 24 123 265 545 532 1 14 1 15 20 876 1262 1094 847 93 106 Table 38. Landsat classification performance/ minimum distance-to- mea ns classifier/ Crawford County test site. LANDSAT CLASSIFICATION COVER TYPE (MAP) Red Pine Jack Pine Pine Mixtures Red Pine — — Jack Pine Pine Swamp Mixtures Conifers Other Total — 321 5858 — — 31 29 381 239 700 6797 42 105 827 735 28 1547 47.5 4127 5053 82.6 — 680 — 784 Other — 858 — 23 Total — 8501 — 1070 Swamp Conifers Percent Correct 68.9 — Percent Correc t 68.7 — 86.2 — 5034 14605 82.9 73.7 107 A cross tabulation of signatures and cover type map categories* as obtained from the maximum likelihood classifier* was also prepared a contingency table (Table 39) and aggregated (Table 40). into The number of correct ly interpreted jack pine and swamp conifer pixels increased* although this increase was partially offset by larger commission and omission errors. were 8 6 . 6 (OE) or 69.0 percent conifers were 49.0 Accuracies for jack pine (CE) and those (OE) or 6 6 . 6 percent (CE). for swamp Overall classification accuracy was 73.7 percent while coniferous forest* 88.2 considered as a single category, was classified with percent accuracy. 4. Line ar-Combination Classifier To test the hypothesis that the magnitude of change in re fl ectivi ty from band 5 to band 6 would provide a measure for discrimi nat ing among the cover types, the test sites were classified using data from only these two bands. Because of the predicted characteristic increase in brightness values from band 5 to band 6 for conifers, compared to the relativ el y low increase or decrease background features, from B V 6 (BV6-BV5) for the value obtained by subtracting BV5 was used to classify the test sites. the Wexford County test site, For the numeric difference between band 5 and band 6 ranged from 0 to 23 (Table 41). The computer was not capable of analyzing negative numbers thus, these were assumed to be zero. As predicted, lower values, 108 Table 39. Two -wa y cross tabulation/ maximum likelihood classifier/ Crawford County test site. Cover Type Map Signature Name PJl PJ2 PJ3 PJ4 PJ 5 PJ6 PJ7 PJ8 PJ9 Q1 Q2 Q3 Q6 Q7 Nl N2 N3 N5 Red Pine Jack Pine 20 613 801 389 130 155 526 946 476 667 42 12 2 4 40 47 38 43 Pine Mixtures 40 83 51 13 16 97 109 58 65 0 1 0 36 129 25 1 1 0 3 5 90 17 438 471 13 11 26 5 0 101 14 2 54 59 36 9 Swamp Conifers 66 118 94 40 24 82 103 44 78 142 176 167 191 56 23 9 3 0 Oth 286 137 41 14 22 85 287 517 646 2 28 1 16 9 775 1315 1024 954 109 Table 40. Landsat classification performance* maximum likelihood classifier, Crawford County test site. LANDSAT CLASSIFICATION COVER TYPE (MAP) Red Pine Red Pine Jack Pine Pine Swamp Mixtures Conifers Other Total Percent Correct — 306 — 48 27 381 — 5932 — 263 651 6846 — 687 — 45 100 832 — 768 — 760 22 1550 49.0 Other — 902 — 25 4069 4996 81.4 Total — 8595 — 1141 4869 14605 — 69.0 — 66.6 83.6 Jack Pine Pine Mixtures Swamp Conifers Percent Correct — 86.6 — 73.7 110 Table 41. Two-wa y cross tabulation/ County test site. B V 6 - B V 5 / Wexford Cover Type Map BV6-BV5 - /0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 Red Pine 185 40 50 68 94 110 116 157 198 272 317 386 495 534 628 594 461 318 190 83 44 16 7 4 Jack Pine Pine Mixtures 136 45 49 70 130 167 227 185 208 176 128 78 45 28 4 37 2 2 0 0 0 0 0 0 0 11 7 18 20 30 50 56 89 121 149 149 184 128 78 46 24 10 2 2 0 0 0 0 Swamp Conifers 8 1 6 11 14 26 36 36 42 52 50 27 21 5 3 0 0 0 0 0 0 0 0 0 Othe 9076 995 884 764 695 567 416 328 224 189 127 87 59 24 13 5 3 3 0 0 0 0 0 0 Ill from 0-5# were associated with background features# intermediate values# pine, from 6 - 8 # were associated with jack and the higher values# pine. By aggregating 9-23# were associated with red the numeric difference values by appropriate cover type classes# was produced (Table 42). to jack pine and "other" for an accuracy of 81.2 percent and for an accuracy of 71.4 percent. "other," and jack pine Jack pine had both large omission errors and commission errors# (CE). pixels# The especi ally including for an acc uracy of 50.0 (OE) or 40.3 percent "other" cat eg or y was fairly well classified# (OE) or 92.9 percent (CE)# 96.1 for an overall classification acc uracy of 81.1 percent. category# a confusion table The red pine class had omissions commissions including pine mixtures# "other" (B V 6 - B V 5 ) Coniferous forest, as a single was classified with 93.4 percent accuracy. As shown in Table 43, the Crawford County test site has a smaller range of values obtained by subtracting BV5 from BV 6 (from 0 to 16). progression conifers# These values also disp lay a similar from background to red pine, different. Again# features# jack pine# swamp although individual ranges were the numeric difference (B V 6 - B V 5 ) values were aggregated from Table 43 to produce a co nti nge ncy table (Table 44). Jack pine was the only class with more commission errors, esp ec iall y to "other" and swamp conifers# than omission errors for an accuracy of 88.4 (OE) or 63.9 112 Table 42. Landsat classification performance, Wexford County test site. BV6-BV5, LANDSAT CLASSIFICATION COVER TYPE (MAP) Red Pine Jack Pine Red Pine 4716 684 — — 407 5807 81. 2 Jack Pine 559 1000 — — 441 2000 50.0 Pine Mixtures 967 284 — — 92 1343 — Swamp Conifers 177 171 — — 49 397 — Other 182 345 — 12981 13508 Total 6601 2484 — Percent Correct 71.4 40. 3 — Pine Swamp Mixtures Conifers Other Total — Percen t Correct 96.1 13970 23055 92.9 81.1 113 Table 43. T w o-w ay cross tabulation/ test site. B V 6 - B V 5 / Crawford County Cover Type Map BV6-BV5 -,0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Red Pine Jack Pine 18 7 9 15 17 36 51 42 39 29 459 250 383 596 725 885 833 699 533 339 158 22 66 15 25 10 11 2 1 0 20 3 1 1 Pine Mixtures 91 26 32 51 65 81 95 104 63 54 33 18 9 5 3 0 0 Swamp Conifers 24 29 47 68 95 140 170 208 233 168 122 67 23 16 4 1 1 Other 3040 540 522 499 492 355 260 200 135 63 28 17 4 2 1 1 0 114 Table 44. Landsat classification performance/ Crawford County test site. BV6-BV5/ LANDSAT CLASSIFICATION COVER TYPE (MAP) Red Pine Jack Pine Pine Swamp Mixtures Conifers Other Total Percent Correct Red Pine 2_ 318 38 22 380 Jack Pine 1 5985 89 695 6770 Pine Mixtures 0 701 29 95 825 Swamp Conifers 2 1456 93 64 1615 5 .7 Other 0 902 9 4104 5015 81.8 Total 5 9362 258 40.0 63.9 36.0 Percent Correct 0. 53 88.4 — 4980 14605 82.4 69.7 115 percent (CE). Overall classification accuracy was 69.7 percent whereas co niferous forest; as a single category/ was classified with 87.8 percent accuracy. Due to the effect of the lowering of (BV6-BV5) at higher absolute brightness values, classified using the (BV6-BV5) nearly 3,000, the test sites were data in conjunction with the absolute band 6 brightness value. combinations of (BV6-BV5) values Since the possible and BV 6 values was very large, the B V 6 values were grouped into classes of ten brightness values each before analysis (Table 45). (BV6-BV5) The values for red pine vary from 21 in the BV 6 range of 11-20 to 5 in the BV 6 range of 101-110. pine is characterized by (BV6-BV5) values only in the BV 6 range of 40 and lower. (B V 6 - B V 5 ) values were aggregated confusion table omission errors, (Table 46). Likewise, jack from 4 to 10, but The BV 6 and from Table 45 to produce a The red pine class had few primarily to "other," for an accuracy of 87.1 percent, but a larger number of commission errors, espec ial ly including pine mixtures and "other," accuracy of 74.1 percent. accurate, The other forest types were less but the "other" class, (CE), was quite accurate accuracy of 84.0 percent. for an 97.0 (OE) or 72.6 percent for an overall classification Coniferous forest, as a single category, was 93.6 percent accurate. The B V 6 , BV6-BV5 values obtained for the Crawford County test site are given in Table 47 and 48. Although 116 Table 45. Two-way cross t a b u l a t i o n , B V 6 , B V 6 - B V 5 , Wexford County test site. Cover Type Map BV 6 0-10 BV6-BV5 -,0 4 5 6 7 8 9 10 11-20 -,0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 21-30 -,0 1 2 3 4 5 Red Pine Jack Pine 0 0 0 0 0 0 0 0 0 1 2 0 0 0 1 0 0 0 2 0 0 3 4" 8 3 1 1 7 13 30 51 45 58 45 37 22 20 , 20 11 10 10 1 10 0 1 0 0 0 0 0 0 1 0 0 1 4 5 5 3 3 3 3 4 6 7 18 26 57 9 1 1 14 14 6 8 3 14 1 Pine Mixtures Swamp Conifers Other 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 2 0 0 0 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 2 9 3 3 4 5 13 17 27 33 15 6 4 1 1 0 0 0 0 0 21 0 0 0 1 2 37 63 64 80 74 3 9 15 24 51 3 1 12 8 4 3 1 0 0 0 0 0 0 0 0 0 0 0 5 4 3 7 20 23 21 1 2 3 3 4 3 3 1 1 1 0 0 0 0 0 0 0 0 0 69 24 25 26 25 30 26 30 37 35 117 Table 45 (cont'd.). 94 135 208 253 324 341 269 178 107 47 10 11 12 13 14 15 16 17 18 19 20 21 22 31-40 21 10 5 -,0 3 1 2 0 0 2 3 4 5 9 15 6 20 7 27 47 75 93 107 164 151 182 172 135 96 60 27 17 3 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 1 1 23 41-50 0 1 2 3 4 5 6 7 8 9 10 50 31 59 75 29 13 30 13 20 112 86 10 11 3 3 55 32 14 3 2 0 0 0 0 0 0 0 0 2 1 1 0 0 0 0 0 0 2 1 1 257 7 3 1 1 0 0 0 0 0 0 7 4 8 20 33 41 53 40 34 31 26 20 6 6 1 0 1 0 0 0 0 0 0 0 2 1 11 4 5 12 12 20 19 33 43 57 61 9 16 28 27 38 22 24 17 9 1 1 0 0 0 0 0 0 1 4 5 10 17 27 25 25 28 34 29 16 9 6 5 1 1 0 0 0 0 0 1 0 2 3 7 8 10 10 16 24 110 3 123 84 6 101 15 18 101 11 13 12 11 7 5 1 1 0 0 0 0 0 0 0 0 0 3 0 4 3 3 5 8 2 3 5 2 79 81 46 58 36 26 17 7 4 1 1 2 0 0 0 0 0 0 406 196 201 194 184 161 107 68 62 38 29 118 Table 45 ( cont' d. ). 11 12 ' 13 14 15 16 17 18 19 20 21 22 23 51-60 -,0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 61-70 -,0 1 2 3 4 5 6 7 8 9 10 11 12 72 58 60 67 45 31 18 13 3 2 1 1 1 3 3 3 0 0 0 0 0 0 0 0 0 0 5 16 16 9 12 12 13 18 31 36 38 37 30 32 33 29 19 15 18 18 16 7 7 7 6 6 8 12 4 3 1 1 16 4 11 20 23 27 16 17 29 23 25 18 14 2 1 2 1 0 0 0 0 0 0 0 0 33 5 4 5 10 6 3 4 0 1 1 0 0 8 13 3 3 1 2 0 0 0 0 0 0 0 2 0 3 3 2 5 10 5 7 3 9 3 6 4 0 0 0 1 0 0 0 0 3 3 1 5 2 4 4 3 2 5 2 0 2 3 3 0 0 0 0 0 0 0 0 0 0 0 20 14 4 3 2 1 0 0 0 0 0 0 0 2 0 1 0 1 2 0 0 2 2 0 0 0 0 0 0 0 0 0 0 0 0 715 195 1 0 0 0 0 1 1 0 0 0 0 0 0 892 171 151 116 201 197 176 149 102 72 36 35 18 18 11 4 3 1 0 0 0 0 0 0 121 76 45 28 30 15 11 7 4 119 Table 45 71-80 (cont'd.). 13 14 15 16 17 18 13 -,0 38 9 1 2 3 4 5 1 12 10 11 12 13 14 15 16 18 8 6 1 1 2 6 8 9 -,0 1 2 40 0 0 0 0 0 0 0 0 4 3 9 3 4 9 2 1 1 0 0 0 1 0 1 0 0 0 0 3 1 1 0 0 2 4 2 2 1 1 2 2 1 1 2 0 0 0 0 0 2 2 1 0 1 21 2 0 0 0 1 0 0 0 0 0 0 0 0 1 1 1 0 0 0 -,0 35 7 9 1 2 11 2 1 1 2 1 1 0 0 2 1 0 3 4 5 6 7 8 9 10 11 13 15 91-100 7 4 18 18 17 17 13 7 17 4 7 4 7 81-90 10 8 3 4 5 6 7 4 11 6 4 5 5 5 4 3 1 7 3 3 8 0 6 8 2 0 1 1 1 1 3 0 2 0 3 0 0 0 0 0 0 2 1 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 747 0 0 0 0 0 0 0 0 0 0 0 0 0 0 703 96 36 32 14 18 0 0 0 0 0 0 0 0 0 781 46 24 15 102 94 87 52 44 17 15 6 3 2 2 1 3 0 0 0 0 11 8 3 1 0 0 1 0 12 4 3 2 1 120 Table 45 (cont'd.)9 10 11 101-110 -,0 1 2 3 4 5 6 7 8 9 111-120 -,0 1 2 3 4 5 121-127 - /0 1 2 3 0 0 0 0 2 0 1 0 1 0 0 0 0 0 0 30 3 6 895 2 2 1 2 0 0 0 0 0 0 1 0 1 0 1 2 0 0 1 0 1 1 0 0 0 0 0 0 0 0 1 0 0 0 0 10 0 1 0 0 0 1 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 1380 16 9 3 2 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2220 5 3 21 12 7 7 2 1 0 0 1 1 1 15 5 2 121 Table 46. Landsat classification performance/ BV6-BV5/ Wexford County test site. BV6/ LANDSAT CLASSIFICATION COVER TYPE (MAP) Red Pine Jack Pine Red Pine 5037 218 30 — 495 5780 87.1 Jack Pine 436 984 43 — 411 1874 52.5 Pine Mixtures 907 246 59 — 108 1320 4.5 Swamp Conifers 149 168 8 — 51 . 376 Other 269 144 2 — 13290 13705 Total 6798 1760 142 — 14355 23055 Percent Correct 74.1 55.9 41.5 Pine Swamp Mixtures Coni fers Other Total 92.6 Percent Correct — 97.2 84.0 122 more categories were obtained in the Crawford County test site# with the exception of jack pine/ lower. percent/ all accuracies are Jack pine was classified with an accuracy of 87.6 considering omission errors only/ or 68.4 percent/ considering commission errors only. Overall classification accuracy was 73.8 percent whereas coniferous forest/ single category/ was 88.9 percent accurate. the linear combination classifier 5. Accuracies from (EV 6 / BV6-BV5) those of all previous attempts for both as a exceed test sites. Boundar y Errors Boundar y errors are defined as mis-classified pixels which occur along the bou ndary between a coniferous forest type and "other." One map from each test site/ with smaller allowable cluster radius/ boundary errors/ clustering was analyzed for including omissions and commissions. Of the 2352 omission errors in the Wexford County s i t e f 1523/ or 64.8 percent/ were bou nd ar y pixels. Similarly/ the 738 commission errors were bou ndary pixels/ 598 of which represents 81.0 percent of the commission errors. Collectively/ 2121 pixels/ or 6 8 . 6 percent/ which were mis- classified in the Wexford County site were bou n da ry errors. In the Crawford County test site/ 971 of the 1518 omission errors were boundary pixels and 1173 of the 1850 commission errors were bou nd ar y pixels. Therefore/ 2144 error pixels/ 123 Table 47. Tw o-way cross t a b u l a t i o n , BV6, County test site. BV6-BV5, Crawford Cover Type Map BV 6 BV6-BV5 Red Pine Jack Pine Pine Mixtures 0-10 8 0 0 0 11-20 2 0 0 0 0 0 0 0 1 0 0 0 0 2 1 2 1 0 0 0 0 1 0 0 0 0 0 0 0 3 4 5 6 7 8 9 10 11 12 13 21-30 -,0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 31-40 — r0 1 2 3 4 5 1 1 1 2 0 1 4 9 10 6 7 2 1 2 0 0 0 0 0 1 6 3 4 9 15 7 21 3 2 2 0 0 0 0 0 4 8 20 1 1 1 6 12 71 91 170 155 131 107 51 15 13 20 8 2 1 0 0 0 2 0 2 1 0 0 0 36 42 85 158 250 319 331 294 16 16 22 1 3 5 14 21 33 42 49 Swamp Conifers Other 1 0 0 0 1 0 0 0 0 1 1 0 0 0 0 4 12 19 24 48 58 51 39 16 7 3 5 9 5 5 5 20 8 33 56 79 92 116 72 54 33 13 18 16 19 27 18 1 11 2 1 1 3 3 14 10 25 34 39 38 8 4 3 0 0 0 0 0 41 32 34 54 76 54 34 43 124 Table 47 (cont'd.). 8 9 10 11 12 13 14 15 16 41-50 -,0 1 2 3 4 5 234 158 71 24 8 8 0 6 4 3 2 1 1 1 0 0 0 0 0 0 0 39 34 9 3 3 1 88 147 185 214 9 3 1 4 7 9 5 11 10 2 11 12 13 14 15 8 5 1 1 6 7 8 51-60 19 11 15 7 6 0 -,0 1 2 3 4 5 6 5 7 3 6 7 8 9 10 11 12 13 14 61-70 6 73 50 61 89 90 97 80 71 55 32 19 10 10 41 31 22 12 2 0 2 0 0 6 B 11 22 35 23 6 3 1 2 0 0 0 97 61 94 110 8 8 6 0 2 0 5 3 127 91 81 60 34 15 7 4 1 1 0 0 2 0 0 1 4 8 2 9 7 9 9 9 8 8 2 8 12 11 5 5 7 4 237 85 96 103 103 92 65 33 26 14 16 17 13 15 5 4 3 1 1 21 5 4 4 19 3 2 2 2 1 1 1 5 3 2 2 1 0 3 102 54 65 73 55 41 43 8 0 1 2 3 4 5 1 2 1 3 7 200 162 117 77 40 14 10 3 0 1 29 20 17 19 23 22 10 10 2 12 10 3 3 0 1 1 1 0 0 20 2 355 9 7 3 112 100 8 6 4 5 3 5 5 2 0 85 65 49 31 125 Table 47 (cont'd.). 7 8 9 10 11 12 13 71-80 -,0 1 2 3 4 5 6 7 8 9 10 11 81-90 -,0 1 2 3 4 5 6 7 8 9 10 91-100 -,0 1 2 3 4 5 6 7 8 9 10 101-110 -,0 1 2 3 4 3 4 1 0 1 0 0 4 0 1 2 3 0 2 2 1 1 0 0 27 16 13 3 1 2 1 82 39 38 42 37 31 11 7 2 7 1 0 1 1 1 0 0 0 0 2 0 0 0 0 0 0 12 8 1 6 1 2 1 416 85 82 56 53 26 14 9 4 4 3 2 2 0 0 0 0 2 1 2 0 1 0 0 0 0 0 0 72 18 20 21 11 12 2 2 2 1 1 2 0 1 0 3 28 4 3 3 3 3 0 0 0 0 0 0 0 0 0 0 2 0 0 0 0 9 9 3 0 1 2 4 19 1 1 0 0 0 2 0 1 1 0 0 0 0 0 0 16 7 1 0 1 0 1 0 0 0 3 0 1 0 0 0 0 0 0 0 7 1 0 1 0 0 1 0 2 0 0 0 0 0 1 0 0 0 0 0 449 76 43 38 27 13 7 5 5 1 0 0 0 0 0 0 1 0 0 0 387 35 42 25 13 1 0 0 0 0 326 24 1 0 8 3 8 3 1 1 12 11 10 126 Table 47 (cont'd.)* 5 6 7 6 111-120 -,0 1 2 3 6 121-127 -,0 1 2 3 6 0 0 0 0 0 0 1 0 0 0 ‘ 0 0 0 0 0 0 6 0 0 0 0 0 7 0 0 0 0 2 0 0 0 0 0 0 0 0 0 301 15 0 0 0 0 0 0 0 0 ■ 0 0 0 0 0 0 0 0 0 0 0 0 426 4 3 2 2 2 1 10 2 2 1 127 Table 48. Landsat classification performance/ BV6/ B V 6 - B V 5 1 Crawford County test site. LANDSAT CLASSIFICATION COVER TYPE (MAP) Red Pine Jack Pine Red Pine 11 312 1 31 23 378 Jack Pine 5 5928 0 219 616 6768 Pine Mixtures 3 686 2_ 38 93 822 Swamp Conifers 2 903 0 608 38 1551 39. 2 Other 1 838 0 16 4231 5086 83.2 Total 22 8667 3 912 50.0 68.4 66.7 66.7 Percent Correct Pine Swamp Mixtures Conifers Other Total Percent Correct 0.03 87.6 0.00 5001 14605 84.6 73.8 128 or 63.8 percent were boun dar y errors in the Crawford County test site. B. Effects of Accuracy Assessment Procedures Errors of omission and commission/ the cover type maps/ as determined were compared with aerial from photographs (Figures 3/4/ and 5) to validate the error and determine the probable cause of m i s - c l a s s i f i c a t i o n . One-fourth of the mis-classified pixels were located on the photography and classified (this was the largest practical sample which could be drawn due to the difficu lty associated with locating individual pixels). For the Wexford County test site/ 67 percent of the error pixels/ as determined cover type map/ were likewise considered aerial photography from the in error from the (Table 49). Table 49. Number of omission and commission errors determined from maps and aerial photography/ Wexford County test site. Aerial Photography Map Omission Omission 189 Commission — Commission — 21 Forest Other — 65 38 — Approximately 60 percent of the omission errors were determined to be boun da ry pixels/ stands of low stocking 20 percent occurred (generally less than 50 percent crown c l o s u r e ) / and 20 percent were unexplained by photo interpretation. in Nea rl y all the commission errors were 129 attributable to either edge pixels low stocking (26 percent). (10 acres) (68 percent) The minimum type size for a map tends to generalize forest boundaries such that over 25 percent of the omission errors were cover types. or areas of in fact other An even greater consequence of the type map generalization was the erroneous listing of a pp rox imat ely 64 percent of commission errors. confusion resulted percent) Most of the commission from either areas of low stocking (61 or edges and small open areas within the forest (32 p e r c e n t ). For the Crawford County test site, the error pixels, as determined were verified as being (Table 50). were only 52 percent of from the cover type map, in error from the aerial photography Over 36 percent of the mapped omission errors found not to be errors by photo interpretation. Table 50. Number of omission and commission errors determined from maps and aerial photography, Crawford County test site. Aerial Photography Map Omission Commission Omission Commission Forest 171 — — 98 241 — — 192 Other Approximately 41 percent of the omission errors were determined to be bou ndary pixels, 28 percent occurred in 130 stands of low stocking, photo interpretation. and 31 percent were unexplained by Errors of commission were fairly evenly divided between edge pixels {38 percent) low stocking (41 percent). Over half the commission errors were classified as forest from photo therefore, and areas of interpretation, not considered to be m i s - c l a s s i f i e d . and Most of these forested areas were either bou nd ar y pixels (41 percent or stands with low, somewhat scattered stocking (25 percent). C. Relative Efficiency of Classifiers To evaluate the relative efficie ncy of the classifiers, with respect to accuracy, the Kappa statistic was utilized to rank the algorithms and to conduct pairwise tests of significance sites, (Table 51 and 52). Comparing the two test Wexford County and Crawford County, indicates that the various algorithms were ranked di ffe ren tly for the two sites. The B V 6 , BV6-BV5 classifier was the most accurate algorithm over both test sites. classification algorithms, maximum likelihood, were, The two supervised minimum distanc e-to -me ans and with one exception, more accurate than any of the unsupervised clustering classifications. Default clusters were re la tiv ely inaccurate, 8 th, whereas level slicing in conjunction with a smaller allowable cluster radius tended BV6-BV5 algorithm, ranked 6 th and to increase accuracies. the "simplest" combination tested, ranked as 6 th or 7th in accuracy. The was All algorithms, with the 131 Table 51. Summary of Landsat classification performance/ Wexford County test site. ALGORITHM ! | ACCURACY OMISSIONS COMMISSIONS OVERALL [ | KHAT RANK Default cluster 91.1 77.7 80.3 0.5951 8 cluster with smaller radius 88.4 82. 2 82. 2 0.6580 7 Level sliced default clusters 89.4 83.0 83.0 0.6779 5 Level sliced cluster with smaller radius 90.0 83.6 83.6 0.6892 2 Minimum distance 84.5 83.1 83.1 0.6864 3 Maximum likelihood 84.4 85.3 83.1 0.6819 4 BV6-BV5 87.7 81.1 81.1 ■0.6643 6 BV 6 / BV6-BV5 85.4 84.0 84.0 0.7091 1 132 Table 52. Summary of Landsat classification performance/ Crawford County test site. ALGORITHM | j ACCURACY O M I S S I O N S C O M M I S S I O N S O VE R AL L | | KHAT RANK Default cluster 79.5 72.9 73.2 .05454 6 Cluster with smaller radius 80.4 73.5 73.5 .05617 4 Level sliced default clusters 79.5 72.9 72.9 .04779 8 Level sliced cluster with smaller radius 79.3 72.9 72.8 .05576 5 Minimum distance 80.0 73.7 73.7 .05617 3 Maximum likelihood 80.4 73.7 73.7 .05627 2 BV6-BV5 73.9 69.7 69.7 .04825 7 B V 6 / BV6-BV5 73.8 73.8 73.8 .05628 1 133 exception of the minimum distance -to-means and the cluster with a smaller allowable cluster radius for the Crawford County test site, had significantly different KHAT values at the 95 percent confidence level. which classifier was used, In addition, regardless of classification of the Wexford County test site was more accurate than the Crawford County test site. While accuracy must be considered a primary measure of a classifier's performance, there are several other attributes which need to be considered for a complete evaluation of efficiency. A summary of the major attributes for the several algorithms, including the number of categories classified, computer execution time, and the need for additional analysis, five classes, is presented in Table 53. or categories, reference source, Although were recognized and mapped as a the algorithms had varying degrees of success in the number of categories which were obtainable. Virtually all the unsupervised clustering algorithms were limited to three categories in either test site. The supervised techniques (minimum distance-to-means and maximum likelihood) produced three categories in the Crawford County test site and four in the Wexford County test site. The linear combination classifier (B V 6 , B V 6 - B V 5 ) came closest to the reference source with four categories over the Wexford County test site and five over the Crawford County test site. 134 In order to compare the amount of time required by each algorithm to cla ssify the data, the elapsed execution time for each algorithm was noted and is expressed as a multiple of the "least" time classifier (Table 53). A wide range of times were noted with the unsupervised clustering algorithms requiring from four and a half to six and a half hours to classify the Wexford County test site, which is 8 to 12 times longer than the fastest classifier (B V 6 - B V 5 ). Supervised clas sifications required over an hour to execute and were 2 to 3 times slower than the fastest classifier. The BV 6 , BV6-BV5 algorithm classified the test site a half hour, slightly (1.19x) in about slower than the fastest time. All of the algorithms available require additional analysis before a classification is obtained. Unsupervised clustering requires post-analysis in order to assign category labels to the clusters. The analysis which utilized level slicing required both a pre-analysis effort, creating a level sliced scene, and a post-analysis effort, assigning category labels to the clusters. Both supervised techniques require that categories be defined through training site selection prior to implementation of the algorithm. The two linear combination algorithms are similar to the clustering techniques in that cat egory labels must be assigned to specific numeric results, on supplementary references. usually based 135 Table 53. Comparison of classification algorithm attributes, NUMBER OF CLASSIFICATION CATEGORIES ALGORITHM Wexford Co. Crawford Co. TIME RATIO* ADDITIONAL ANALYSIS PREPOST B V 6 1 BV6-BV5 4 5 1.19 x BV6-BV5 3 4 1.00 x Maximum 1 ikelihood 3.32 Min imum distance 4 3 2.48 Default cluster 2 3 11.48 Level sliced default cluster 8.55 Cluster with smaller radius 12.71 Level sliced cluster with smaller radius 10.16 X ♦Execution on the ERDAS microcomputer expressed as a multiple of the time for the BV6-BV5 algorithm {actual time was thirty-one minutes for the Wexford County test site). 136 For relatively small sites, size, an equal amount of time approximately a township in (four to six hours) is required for either pre-analysis and/or post-analysis classification assignments. With larger test sites, especially those approaching full scene size, training {a pre-analysis technique) (eight to twelve hours) labels to categories will take supervised will require more time than the time required to assign (a post-classification from six to eight hours). technique which CHAPTER V SUMMARY AND CONCLUSIONS This study has evaluated spectral the use of Landsat multi- scanner digital data for classifying and mapping coniferous forest cover types. All analyses were conducted on a Landsat scene obtained on February 26, 1979 which is centered in the north-central Lower Peninsula of Michigan. The scene recorded a landscape under a ubiquitous snow cover with coniferous forests providing the only green-foliage reflectances in the entire scene. Two test sites were chosen/ one in Wexford County and the other in Crawford County, to be representative of areas now supporting large acreages of conifers. were prepared Cover type maps of the two test sites from aerial photography, digitized, rectified to match the Landsat data files. and then Subsequent classifications from the Landsat data were compared with these "reference" files to produce error matrices. Several standard digital analysis techniques (i.e. algorithms available on the ERDAS micro-computer; unsupervised clustering, maximum likelihood) sites. In addition, minimum distance-to-means, and were utilized to classify the test the effect of varying the values of input parameters on the accuracy of the unsupervised clustering algorithm was evaluated. employed with unsupervised clustering 137 Level slicing was also in an effort to 138 minimize the effect of a large number of non-forest clusters. A spectral response curve model was developed from analysis of the multispectral reflectance patterns exhibited by the coniferous cover types and the background The predicted brightness values features. from the model were utilized to construct a 1 inear-combination classifier which was also tested for classification accuracy. In order to evaluate the effectiveness of the cover type maps as verification sources, tests were conducted using aerial photography as the "ground truth." Dis cr e­ pancies were noted between the two methods and possible causes investigated. Each of the classification with respect to its overall techniques was evaluated classification accuracy, the magnitude and source of errors, the ranking and significance based upon the kappa statistic, the number of categories obtainable, execution time required, and the need for additional analysis. The major findings of this study can be summarized as follows: 1. Unsupervised clustering, provided using default parameters, the least accurate (80.3 percent) classification of the Wexford County test site and was ranked sixth of 8 for the Crawford County test site. This algorithm produced a large number of errors/ both of omission and commission/ and was especially error prone where stands were small and/or irregularly spaced. The only input variable which con sistently affected the classification clustering performance of the technique was the maximum allowable cluster radius. The reduction of this variable from seven to three digital counts increased the accuracy from 80.3 to 82.2 percent and from 73.2 to 73.5 percent for the Wexford County and Crawford County test sites/ respectively. Level slicing of the scene prior to clustering increased the accuracy for the Wexford County test site/ but had the opposite effect County test site. for the Crawford Clustering level sliced scenes in conjunction with a smaller allowable cluster radius improved accuracies for both test sites. With one exception/ the supervised classification algorithms/ minimum distance-to-means and maximum likelihood/ had higher overall classification accuracies than did the unsupervised clustering algorithms. The minimum distance-to-means algorithm was more accurate than the maximum likelihood algorithm over the Wexford County test site but the opposite was true for the Crawford County test site. More errors of omission occurred, commission errors, compared to and were largely attributable to lightly stocked stands, (<50% crown closure). A spectral response curve model was developed which could predict brightness values from various mixtures of conifers and background features. The predicted brightness values from stands containing a mixture of conifers and background features demo nstrated that the magnitude of change in refle cti vity from band 5 to band 6 provides the most consistent measure for discriminating among the cover types. Even a simplistic version of a two-band linearcombination classifier (BV6-BV5) was more accurate than either clustering with default parameters or clustering with a smaller allowable cluster radius for the Wexford County test site. Crawford County test site, Over the this algorithm was more accurate than clustering of a level sliced scene. A slightly more sophisticated 1 inear-combination classifrer which uses the (BV6-BV5) data in conjunction with the absolute band 6 brightness value (i.e. B V 6 , BV6-BV5) accurate classifications, produced the most 84.0 and 73.8 percent for Wexford and Crawford Counties, respectively. Post-classif ica tion analysis of aerial photography indicated that approximately 33 percent of the "errors" in Wexford County and 48 percent of the "errors" in Crawford County were attributable map generalizations. to Approxim ate ly half the errors were attributed to bou ndary pixels, another 40 percent were associated with thinly stocked stands. The remaining errors were caused by small openings in the forest (below the minimum map size but larger than the IFOV of the Landsat MSS). The number of mappable categories varied among various algorithms. Unsu pervised clustering techniques produced, at most, Supervised categories, three categories. techniques produced from three to four while the linear combination classifiers produced from three to five. Execution time varied considerably. clustering was the slowest, to six and a half hours; were the intermediate, from four and a half supervised techniques from one and a quarter and three quarter hours; classifiers were Unsupervised to one and linear-combination the fastest, about one half hour. All of the algorithms tested require additional analysis before classification Except for the level is complete. sliced analysis, which 142 requires both pre- and p o s t - a n a l y s i s , each algorithm requires one additional step to assign ca tegories to numeric results or to specify training site data. 13. The relative performance of the algorithms differed between the two test sites such that different rankings were allocated to the algorithms by site. 14. Overall classification ac curacy was significantly differe nt between the two test sites. contributing factors appeared The major to be the blocky plantation pattern in Wexford County compared to the scattered/ heterogenous forest cover Crawford County. classification/ in Even the least accurate 80.3 percent/ for the Wexford Coun ty test site was superior to the most accurate classification/ 73.8 percent/ for the Crawford County test site. 15. Digital classification accurate than visual enhanced/ spring techniques were more interpretation of computer imagery (72.7 percent) Crawford County test site/ over the but were less accurate than results from the Wexford County test site . (84.3 p e r c e n t ) . With respect to the above findings/ certain conclusions can be drawn on the appropriate use of Landsat multi- 143 spectral scanner data in forest resource inventory systems under Lake States conditions. procedures can While digital classification identify coniferous forests with acceptable accuracy (approximately 90 percent)/ accuracies are highly variable. individual cover type Accuracies range from over 90 percent to under 10 percent and also vary by site. Forest cover type maps/ as currently compiled/ include de lin eations of forest cover types and stand size and stocking classifications which cannot be derived directly from satellite data. Thus/ Landsat multispectral scanner data cannot entirely replace traditional/ photo-derived forest inventories. types of assessments/ For more generalized Landsat data is probably a sufficient/ stand­ alone information source. The greatest utility for Landsat data is likely to occur in a comprehensive inventory system utilizing multi­ stage sampling. The availability of remotel y sensed data at several scales provides an efficient sampling very large areas. technique over A large number of fast/ relatively inexpensive measurements can be obtained from the satellite data and correlated with samples from pro gressively higherresolution data sources/ such as aerial eventu all y ground plots. photography and Variable probability sampling/ * with the probability of sample selection proportional to the sizes (or acreages) formulated estimated from the previous stage/ from additional are information available at each 144 stage. At the last stage, measurements are collected in the field and projected back through the sampling formula to obtain estimates for the entire area. This technique is especially suitable to large area inventories such as the Forest Inventory and Analysis for the entire state conducted by the U.S. (1980) Forest Service. The last analysis of Michigan utilized aerial photography as the first level of sampling. A total of 176,976 1-acre plots were classified from the photography. A sample of these plots classified stereos cop ically by forest type, class, and density, (83,103) were stand-size and finally 13,991 of these points were measured on the ground. Using Landsat data to stratify forest land as a first level of a multi-stage sample would provide more accurate survey data, with a smaller sample size. or similar accuracies In addition, the Landsat classification would provide a spatial component to the distribution of forest cover types unobtainable from current Forest Inventory and Analysis procedures. Considering the level and accuracy of information obtainable, the Landsat system is extremely efficient. One Landsat scene covers 13,225 square miles and would require ap proxima tel y 5,000 aerial photographs (at a scale of 1:15,840 with 60 percent endlap and 30 percent sidelap) cover the same area. to Computer compatible tapes for a single 145 Landsat scene cost $660 compared to $150,0001 for the acquisition of medium-scale aerial photography. Although the minimum configuration of a computer system to process the Landsat data is app roximately $24,000, compared to $2,000 for photointerpretation equipment, a single scene could be processed within several days compared to several months to interpret aerial photography for an equivalent area. The decision to utilize satellite data or aerial photography will obviously depend upon an analysis of both information requirements and ‘the associated costs. The high temporal frequency of Landsat data acquisi­ tion, an 18-day repetitive acquisition cycle, could also be exploited for inventory updating requirements. and high-altitude aerial photography could effective technique Landsat data provide a cost- for updating the state-wide Forest Inventory and Analysis. A multi-stage sub-sample of plots from the previous inventory would be utilized to derive "change coefficients" to update acreages, growth projections to a mid-cycle point* also been suggested as a source current use inventory. volumes, and Landsat data have for updating the state-wide The advantages of using Landsat are that on ly land use changes, not an initial inventory, would need to be identified and that the digital nature of the data could possibly be used to automatic ally update current ^Approximate cost for the acquisition of 1:15,840, black-and-white infrared aerial photography based upon a cost of $19.20 per flight line mile. 146 computer files. In addition/ the Landsat system might provide data for monitoring changes in forest areas over short time-frame events (e.g. forest fires or defoliation due to insects or disease). Although current capabilities of processing Landsat data can provide valuable inputs into forest resource assessments, further research and newer satellite systems should be considered. For example, since the linear combination classifier is based upon a spectral response curve model which integrates the spatial proportion of conifer versus background in the I F O V , it may also provide a measure of stocking or density. Further research should investigate this re lat ionship and "automating" broad-area addition, its potential for forest stand classification. characteristic response curves should be investigated from other seasons spectral response curve model to test the val idity of the for possible application classifying and mapping deciduous Several new systems, to forest cover types. including the Thematic Mapper on board Landsat 4 and 5 and the French SPOT satellite, increased spatial resolution tively) In (30 and 10 meters, compared to the multispectral scanner. offer re sp e c ­ Although increased spatial resolution should decrease the effects of boundary pixels, the smaller IFOV might be problematic in areas of dispersed forest cover such as encountered in the Crawford County test site. The full ram if ications of 147 increased spatial resolution on overall classification accuracy would need to be fully investigated. spectral and radiometric resolution Increased from the Thematic Mapper has the potential of improving discrimination among similar cover types (e.g. species of pines) and should also be investigated. Ecological c o n s i d e r a t i o n s , especi all y the effect of site on the choice and performance of various classification schemes/ need to be more fully assessed. Both overall classification accuracy and the relative performance of the algorithms tested in this study were significantly different between the two sites. Signature extension does not appear to be valid across an area the size of the northern lower Peninsula. Therefore/ stratification of the scene, along major landform units, should be tested as a possible mechanism for allocating individual classification techniques. possibly APPENDICES APPENDIX A THE LANDSAT SYSTEM* The launch of the first Earth Resources Technology Satellite (ERTS-1) on July 23/ 1972 marked the beginning of a program of remote sensing from space. in the late 1960's, Initially conceived this program was designed to demonstrate the feasibility of rem otely sensing earth resources from unstaffed satellites. Mission requirements were developed by scientists in the National Aeronautics and Space Adminis­ tration and the U.S. Department of Interior. the acquisition of medium resolution/ from systematic/ local time. included multispectral data repetitive observations In addition/ These taken at a constant both photographic and digital data were to be produced and made available to interested users. Both the satellites and the program were renamed Landsat (for land satellite) 22/ 1975. operational Although with the launch of Landsat-2 on January the satellites were designed with an lifespan of one year, these original expectations. in January/ 1978/ they have greatly exceeded Landsat-1/ which was retired acquired more than 270,000 scenes of *The material in this section refers specifically to the first three satellites in the Landsat program. Although the multispectral scanner on board Landsats 4 and 5 is nearly identical to the earlier ones, the platforms and orbits have been changed. In addition, Landsats 4 and 5 are equipped with a seven channel multispectral scanner known as the thematic mapper. 148 149 portions of the Earth during its five and half year operation. Landsat-2 acquired 185/105 scenes from January/ 1975 to July/ 1983 (nearly 8.5 years) acquired 324/655 scenes from March/ 1983. Currently/ multispectral obtained from Landsat-4 (launched March/ A. while Landsat-3 1978 to September/ scanner imagery is being (launched July/ 1982) and Landsat-5 1984). Spacecraft and Orbital Characteristics The actual vehicle for Landsats 1/2/ and 3 is a re ­ configured Nimbus weather satellite (Figure A-l). These butterfly-shaped satellites weigh 959 kg (2100 lb)/ are 3 meters (10 feet) high by 1.5 meters meter (13 feet) wide solar panels. (5 feet) wide with 4 These satellites were launched by a Thor-Delta rocket from Vandenberg Air Force Base in California. The satellites were placed orbit at a nominal A-2). altitude of 917 km (570 miles) This orbit is sun-synchronous/ angle between the sun, (37.5 degrees)/ inclined 99 degrees (Figure A-2) (Figure and the To maintain this constant the orbital plane of the satellite is (measured clockwise from the equator) so that it rotates at a rate equivalent rate of the earth about the sun configuration circular which means that the the center of the earth/ satellite are held constant. angle into a near-polar/ (Figure A - 3 ). to the This insures that the spacecraft will cross over the same area of the Earth at a constant local time/ thereby 150 M SOLAR m PANEL MUL T I S P E C T R A L SCANNER RETURN VIDICON Figure A - l . BEAM CAMERAS Landsat obs ervatory configuration (adapted from Landsat Data Users H a n d b o o k / U.S. Geolo ­ gic al“"Survey"! 1979 ) . 151 LANDSAT AT 12:30 NOON LOCAL TIME EARTH ROTATION EQUATORIAL PLANE LANDSAT ^ AT 9:42 AM LOCAL TIME Figure A-2. Inclination of Landsat orbit to maintain sun sy nchronous orbit (from Taranik, 1978). OHMT H A JU AOTATU AT .................. M W I t TWt M A D M T t O f THE (AATM AICUT THE |(JM f un Figure A-3. Sun synchronous Landsat orbit (from Landsat Data Users Handbook/ U.S. Geoloqical Survey, 1979). 152 creating repeatable (about 9:30 a.m. illumination conditions. local time) A mid-morning overpass time was chosen as providing neither excessi vel y long shadows nor shadowless conditions while avoiding the tendency of afternoon cloud bui ldu p over terrestrial areas. The Sun's rays strike the Earth at different angles during different times of the year and also vary by latitude. At 45° north latitude in the northern Lower Peninsula) (located the solar elevation angle changes from 60° in June to 18° in December (Figure A-4), thereby creating changing illumination conditions seasonally. the azimuth of solar illumination will change In addition/ seasonally providing different directions solar illumination (Figure A-5). It requires 103 minutes for the satellite one orbit about the Earth. Thus, the sun's illumination to the westward progress of (Figure A- 6 ). completes a pass over Michigan 14 orbits. beneath the successive orbits will be displaced westward at a rate equivalent Montana. to complete During a single orbit the Earth will have rotated 2,760 km (at the Equator) satellite. from After the satellite the next pass would be over After 24 hours the satellite will have completed The westward progression of the orbit will be such that the next pass will be one orbit pass to the west of the first orbital pass (Figure A-7). satellite will have completed 251 orbits. After 18 days the The next orbital 153 JUNE D EC TANGENT SUBSATELLITE EARTH "PLANE POIN T ELLIPSOID Figure A-4. Sun illumination relationships for 45° north latitude (Adapted from Landsat Data Users Handbook/ U.S. Geological Survey/ 1979). 154 t* SUMMER 157* AZ WINTER I f i r AZ SUMMER 120* AZ Figure A-5 Variations in azimuth of solar illumination with season and latitude (from Taranik, 1978)- 155 Figure A- 6 . IIP 1IS 160 138 120 100 Orbital coverage characteristics of Landsat {from Landsat Data Users H a n d b o o k s U.S. Geological Survey/ 1979) . 00 TS M 40 30 10 0 10 30 40 00 70 00 IQS 120 130 150 105 1111 G 0LD 3T0K [REPEATS EVERY I I DAYS! * IIP 100 ISO 135 120 10B 90 Figure A-7. 7P M O M IS P 18 30 45 80 70 90 IBS 120 135 150 IIS IIP Landsat orbital tracks for one day of coverage (from Landsat Data Users H a n d b o o k s U.S. Geological Survey, 1979). 156 pass will then coincide with the first pass producing repetitive coverage on an 18-day cycle. The orbital configuration of the satellite reauires acquisition from seven different passes to provide complete coverage of the state of Michigan. Since data acauired within an orbit are segmented into scenes of about 185 km along track, it requires 23 such scenes to cover Michigan. This segmentation process is applied such that consistent scene centering is accomplished with the resulting scenes indexed by a world wide system of paths and rows A - 8 and A-9). (Figures Scenes produced over Michigan will have approximately 40% sidelap between adjacent paths with an arbitrary 1 0 % endlap created single pass (i.e. produced from scenes acquired within a a small amount of redundant data will be for two scenes segmented from a single orbital pass). B. Landsat 1 and 2 Sensor Systems The payloads for Landsats 1 and 2 were identical and included two imaging instruments', a return beam vidicon (RBV) camera system and a multispectral scanner (MSS). The RBV system was designed to acquire high-resolution television-like images of the Earth. The system consists of three cameras which are aligned to view the same 185 km by 185 km ground area (which coincides with the ground area coverage of the MSS) utilized to obtain (Figure A-10). The three cameras are images simultane ous ly in three different 157 PATH 124 ROW 37 ACTUAL SCENE CENTERS OF REPETITIVE COVERAGE ROW 38 * .' NOMINAL CENTER POINT Figure A - 8 . Relation shi p of actual scene centers to nominal center points for Landsat scenes acquired on a repetitive basis (from Landsat Data Users Handbook, U.S. Geological Survey, 1979) . 158 2B 27 26 ZS 23 2« 22 21 2P W(B. 30M ic h , / >----/— lf— /----- /----- H V «»• / ) O h io Ind. L andsat ■ Landsat notlnsl Figure A-9 Imagery M ic h i g a n lm a g * r y in»a9 e center tocations. 159 THREE RBV CAMERAS MOUNTED IN SPACECRAFT IIS km X I l f km HDO am X 100 nm) D IR E C T IO N OF F L IG H T Figure A-10. RBV scanning pattern (from Landsat Data Users H a n d b o o k / U.S. Geological Survey, 1979). 160 broad spectral bands. solar radiation Camera 1 (Band 1) measures reflected from 0.475 to 0.575 urn (visible, yellow- green), camera 2 (Band 2) 0.580 to 0.680 urn (visible, green-red) and camera 3 (Band 3) 0.690 to 0.830 um (visible red and reflected infrared) (Figure A-ll). Images are generated by the RBV system by shuttering the cameras and storing the resulting image on the photosensitive surface of the camera tube. video signal This surface is then scanned to produce a for subsequent transmission to a ground receiving station. RBV data are available as computer compatible tapes or as film images generated from the digital data. Because of earl y problems with the tape recorders on Landsats 1 and 2 , only a limited quantity of RBV data has been collected. Contrary to pre-launch expectations, become the primary data collection Landsats 1 and 2. the MSS system has instrument utilized on The multispectral scanner is a line scanning device which utilizes an oscillating mirror to scan a 185 km swath perpendicular to the satellite's path. Each active scan produced by the mirror sweep scans six lines simultaneously, collecting data (Figures A-12 and A— 13). in four wavel eng th bands The forward motion of the satellite during the mirror retrace period is such to position the next six scan lines immediately below the last six lines thus providing a continuous scan of the Earth beneath the satellite. Because 161 »< oa IM M IM *AVCUHTH Figure fl-11. m m 61 SI [IM Spectral response/ RBV three-camera system (from Landsat Data Users H a n d b o o k / U.S. Geological S u r v e y , 1979). 162 MOTf; A C t W U U I I I l W i n TO EAST FIELD OF V i m * I I U O E B R I I t MATH IllHUCM rtM S MIT Figure A - 1 2 . Multispectral scanning arrangement (from Landsat Data Users H a n d b o o k , U.S. Geological Survey, 1979). t f A C l CRAFT VELOCITY VECTOR W J ttll J i l l M R W IDTH “ J 1 I I 1 | I C fiftV O K TE TOTAL A R E A ICAR F O R A N tB A N O FOAM ED BY REPEATED H I R E « R BA RD SW EEP! P fA A C U V C JW H R O ft CYCLE LINE Figure A-13. Ground scan pattern for a single MSS detector (from Landsat Data Users H a n d b o o k , U.S. Geological Survey/ 1979). 163 of the continuous satellite motion along track during a scan/ the scan lines are not perpendicular to the orbital path. This opti cal-mechanical scanning process and simultaneous rotation of the earth beneath the satellite produce images which are parallelograms/ not squares (RBV images are square since they are acquired nearly instan­ taneously) . The MSS on Landsats 1 and 2 acquire data wavele ngt h bands; Band 4 detects reflected solar radiation from 0.5 to 0.6 urn (visible/ green)/ (visible/ red)/ A-16, Band 5 0.6 to 0.7 urn Band 6 0.7 to 0.8 urn (reflected infrared), and Band 7 0.8 to 1.1 um (reflected A-15, in four infrared) (Figure A-14, and A-17). The scan mirror of the MSS reflects radiation coming from the surface of the earth (and its atmosphere) detectors through fiber optic bundles. to permit only certain wavelengths Each detector produces a voltage, which is related detector. onto the Filters are utilized to strike the detectors. from zero to four volts, to the amount of radiation that strikes the In order to produce individual area measurements, the output voltage from each detector is sampled during each active scan (Figure A-18). Individual measurements are taken from a ground area of app roximately 76 m by 76 m, instantaneous field of vi ew (IFOV). such that, ships, The sampling rate the is in order to maintain proper spatial rel atio n­ these measur eme nts are formatted as if they 00427-077 023 030 Figure i-14. Landsat bl ac k-a nd- white band 4, E-30556-15460. scene 00427-078 Figure A-15. Landsat black-an d- wh it e band 5, scene E-30556-15460. 165 004Z7-079 H-023 030 Figure A-16. Landsat black-and-white band 6 / scene E-30556-15460. 00427-080 Figure A-17. Landsat black-and-white band 7, scene E-30556-15460. 166 0 ) Ul > m 9 ) tu X. t~ X h 2 § U 16 O 0 ° Figure A-l WATER 400 600 1200 1600 NUMBER S A M P L E S PER 2000 2400 2000 1 6 6 . 2 KM LI NE Rela tionship between voltage and digital count for a hypothetical scan line {from Taranik, 1978). 167 were taken (Slater, from an area of 58 m by 76 m (Figure A-19) 1979). This latter area is called a Landsat pixel (picture element). The voltages produced by each detector for each pixel are converted from an analog signal digital numbers, form by means of a multiplexer. from 0 to 63, to a The resulting are called brightness values (BV) and are dir ectly related to the amount of solar radiation reflected from the surface of the Earth for a specific wavelength band. These data are transmitted receiving station where they are reformatted to a ground into computer compatible tapes and converted into image products. C. Landsat 3 Sensor Systems Based on the experience from sensor operation and subsequent analysis of data from Landsats 1 and 2, two system changes were made on Landsat 3. The three-camera, m u l t i s p e c t r a l , RBV system was replaced by a two-camera, single spectral response, system. Both cameras have the same broad-band spectral sensitivity, from 0.51 to 0.75 urn (green to near infrared). The cameras are configured to produce side-by-side pictures a p pr ox i­ mately 99 km (62 miles) 185 km (115 miles) on a side, thus, swath width as the MSS they cover the same (Figure A-20). Whenever two adjacent series of exposures are made, resulting four images correspond (Figure A-21). the to a single MSS scene To produce this image format, the focal 168 SAMPLING INTERVAL OF MSS ENEROV MEASUREMENT MADE F R O M A B 7 T S S O U A R E M ETE R AREA Figure A-19. FORMATTED TO LANDSAT PICTURE ELEMENT (PIXEL) 4 4 0 1 SQUARE METER P AREA Formation of the MSS picture element from Taranik, 1978). (adapted 169 I■OUNTICH* TmanJeMtfl itt C ttM F f Figure A-20. Scanning pattern of the RBVs on Landsat 3 (from Landsat Data Users H a n d b o o k / U.S. Geological Survey/ 1^7$). 170 GROUND P A T H O F S A T E L L I T E Jl 16.0 Jm , T Y P I C A L IMAGE FORMA T C E N T E R Figure A-21. 0 W orld R e feren ce S ystem nom inal scene c e n te r. T h e actu al lo c a tio n o f scene w ith re sp e c t to th e n o m in al scene c e n te r varies w ith sp a c e c ra ft a ttitu d e . • S cene c e n te r ; m u tu a l in te rse c tio n o f th e fo u r subscenes. Format of a fully processed Landsat 3 RBV scene showing the location of subscenes A /B r C / and D (from Landsat Data Users H a n d b o o k , U.S. Geological Survey, 1979). 171 lengths of the cameras were d o u b l e d t thereby increasing ground resolution to a pp rox imat ely 30 m (Figure A-22). RBV data from Landsat 3 have been primarily used for ca r t o ­ graphic projects requiring increased spatial resolutions. A fifth channel Landsat 3. (Band 8 ) was added to the MSS on This is a thermal channel which detects emitted radiation from 10.4 to 12.6 urn, the resulting quantized values are therefore related to apparent temperature. The IFOV of the thermal detectors is 237 meters square resulting in pixels that are appr oxi mately three times as large as those in bands 4/ 5/ 6 / and 7 (Figure A-23). This channel failed shortly after the launch of Landsat 3 and very few scenes were acquired. D. Landsat MSS Imagery Characteristics The geographical coverage of a single Landsat MSS scene is approx imately 185 km (115 miles) on a side. The geometry of a Landsat image is shown in Figure A-24 and the correspondence between an image and computer compatible tapes is shown in Figure A-25. The annotation block on Landsat imagery provides a summary of sensor operation/proce ssi ng information: listed in Figure A-26. the various codes are A comparison of the coverage of a single Landsat scene with conventional aerial photography (assuming a 9-by-9 inch format/ a 6 -inch focal length lens/ and 60% overlap with 30% sidelap) is given in Table A-l. 172 l ;*■ R s iil Figure A-22. Landsat RBV subscene A/ E - 3 0 0 5 2 - 1 5 4 6 1 . 173 Figure A-23. Landsat black-and-white band 8 / E-30052-15462. 174 IHfOOO00 i UNES/SCAN 555$5M55555555555555555555 BOUNDARY / OF COVERAGE' X = 32*0 COLUMNS 30 MINUTES LONGITUDE AT LATITUDE 40*N WOOO 00 I §1 Figure A-24. Landsat image coverage. Latitude and longitude lines drawn for an area in Colorado (from Taranik, 1978). 175 Cl.11 Figure A-25. Correspondence between a Landsat com puter compatible tapes. image and 176 O r ig in a l a n n o t a t i o n In u s e f r o m J u l y 2 3 , 1 9 7 2 t o F e b . 1 7 , 1 9 7 7 1 e a 'tlittn or jims 71 i i tOUHMrqitMt 4 i * |f*0tn4*7HC1t « i 9 1 e 34*7110 ■*)« MTSMtf 9*1471 fOUNMTIS n*v ism i Sim i k W A i g n tS4.in«.A 9 N I OltMMrt* •M. t 0in«MTI 9 1 HiikMrminni Ml* tSTI <*04M»0U>'I 7 1 ------- o ■tv I ■19 ) MtfMurminu <0 W MfOWWH}. » 2 s. Cfiareosr Poartono 01*00, gf JUN 730er.mendi*Wvaarof a. Character PoaMdne oo»2&. CWWW 1TS.1I tengtfudootmeeonMrafOioRBV an* M U image W irm « «ndlnatad in degree* i w rrarMssm o b end M V japnot P»**r m are i WMCIv gewnetrte dWf ISVW ealonston of mo specscr*! year sm lotho eanh's c. C M n d o PoaiBon* »M3. N KW-WW11S-H UWudeeodloneOude W Iherie* |M ifW M tM l m d l tftt W l T l surfsoe of s psfpsmSouisr moo from the *peo**Mt to wo eoflh sMosowi to inocoad in dsqrsss «no rranutoa. Tho MSSA Oapso* & Characmr Poawono 4 U i RBV t OXA Crwofoi to tttagroup 0 1 son* plated tract from m o SI.II M V Ihurtor Duration Codo *‘XA" moons ms utuwr forCamara iwoo o«i for 4.0 mUltoeeonds* C l a w 3 lor 44 m s*. and Camora 3 for 14 ms. ~ t § r *oUd inwcato mo ttuoar swing forC amoral m 11 ms Camara 3 is 1.4 mo and Camara 3* Time W| t* u « > «« * 4 1 M SS t MrmtatHn tuniijetiaii bseo Code Comoro t Camera 1 Canera A L4 4.0 4* B 0.4 12 34 iul C LO LB tIO 110 110 D ito 1« C TL0 9344 Aeartm CerroOiOft die for m i misoeei 4341 Sonoor * w NOFP I Charsoar Poortion 14 Ground rocording station. 0* OWdston* A-Aisma. N^fTTP. a. Charset* Poesons M l T4U-P.1R imago isft* so*. Chsractsr position 13 metctts nrpootproa N ■ Normal A e Abnormal P ■ prewaed orbit 0 • oaiwova or boat Chsractsr peartttn 17 ma c*** moos of MSI signal procaaeng _ pnerto vananwHw ' rgV to ground staddA. W * 1 ■ snsar mode dockmso from tvs fMrtt al tho 2 ■ cdmpreeaed mods Piwoof M V mpoouro ornWoowt of U M frame is tpeansd io Charsoar peateen U moieawa 1da — gam XK* gam f. Chsraosr P a n ana fAfl, V • lowgam {forBondssandB 1H*1234»A ofty «Neh hot* s commandabia genopdoiii Charsetsr Poawona Wit. toaoocrad nooding to r m h. Character Poarilons M L degree, maaaursd clockses irem NASA CRTS. true North, itio art** pom plus scooscrWt van. Jdanttflss the Agency and Wo Maoooig resets to an knags is o w n tetrerdtho snag* eimore* s Charao SUN CL 31AZB II Bun Angies — aune NOTBBt 0 T H I L I T T i n i THKOUQH “I" m f m TO P A M M A P H B IN THM OOCUM M T THAT aXMJUN T M ANNOTATION BLOCK. C H A M C T M -T O am O N * IN T N I ANNOTATION •LOCK. Figure A-26. Chveresr PoeOsrs 0.TI Pour dipt orott rwokiflon. Nov. *000i" staffswtn in* o a aocsn* ding nodo Isouth to norm squalor gosengi aftor launch. Charaasr poodiorf IS motcaisa apnamans oa« used to compute plated book from the wavr ® IQIWHfftHlH •ft O 0 — Apart** cerrocdon "out" — ttank -— I. Charocssr Poardono W 1 14 (.TM3.iag32.lS Promo idanaaeabon Hum oar — seen unsgsor framefMosursoue ifptaoiion nunoar for. mb * I^UOD^eiMMfl^CPA B — Encoded Project tdarsMsr A — Lands* Mission 1 - Lane** t 2 • Lands* 2 OOO — Oar maneer nfaet* to launen « vme ef oossrta non. rot — Hour * ttme of o m o h OOfL M M - Minutestsms ofobserve bon 3 — Tans of seconds * tune of obsevwn. B — NDPP idansfteason Coda (RBV: T.li MSB 4. 9. 4 7% C — Blank for earth k m o s s Elmer I. t. or a for RBV rsotvnome caabrsoon im* agae mdiciHng lews* io mghssl osposurs lavs* PR — Pegsnarabon number, in* dlc**o insr+4-imofs wdso taps m His event of s manureson uithe poor rim. m M C T N A L IMNTITICATION C O O K M U M H M ANK a T A O M M O TO n m K T IM H TinC A TIO N O F TH1 •FBOTNAL IMA Q I C U M O TO MAKI A COLON C O M F O V T A THANAFAMNCY. © COMFNKMKD DATA WILL M O IC O M M IB M O OURMO F N O C aiaiN O . Landsat annotation block explanation. 177 Annotation In use for NDPF-produced Imagery from Feb. 18, 1977 to Jan. 15, 1978 HU Annotation in use (or NDPP-produced imagery from Jan. 16, 197Q to mid-1978 nu TO DM* CMMACTVH Annotation in use for tPF/EDIPS-produced Imagery after mld-1978 • * • 4 * ftaaacra ftiSMvanatnaaa anaains imatiutwJi w w i n i n r ' i H a n e 1 VlnuVwTira i t M u n /— V a. Character Pownonc 01-04. (4/«w ITJUN T W Day.montfi 91 maaraati longrtudo i tutcantor o*iho RBV «id M W m i a m format is indicated ut d* . For NORF-produeed im* am 40 41 Character position 14 •now* ma rype of MSS • Character petition 71 datinaa ma proracaone ■V - Lambert protection “T - Polar Stereo protec­ tion T-r ■ Haona Ob*awe pro* jecbon -y * Saac* Obkpifo Blana 92*94 And row idereitor, The 203 d«nnfJcatttn Tho “V * UTu groraction " W • H m * m psrtpoawo *'tr >n»ceioe dtreet Character position so mdicataa me rosampung Igornnmi “C* - cubia ■t t ■ neeraat na^noor d«ta d ayad Dick from ina tatatiiia WOVT r N VNU-B3/W11M3U Nominal litituda and 41 t, 92*11 Sensor soecirst band or U m w 3 RBV tueecone iftg. AF*11 (*“ — *97*91 RBV 9iuRsr Ouraaon Cad* Far Uncut 1 AndZKtatteeameaa annotation m u u onortoFeb. tB.1479. For undset 3 “tu­ rntana tna camera woe aal far 2.4 meiaeconds. “XI* in. diesiee me eamara wee mi for 40 mi. *xc* maana as ma, **CT maana 00 me., and '1" fftana 1 2,0 ISO nm t 90 nm) TIS km 1 170 km “T Far MSB irmgw RBV imaoee: "T - ii4 km a 144 km (100 mi 1 lOO nm| T - I U km 1 124 km *13* •nmeateo area data p r a y to d m a from me setoiiiie WBVT r 0 r1 n0 «*10 •1 •4 o" Character poekton II in* dieatat in* (roe or epnemerte data used 10 compute me image comer ■r « txeocbve "tP * osflrreveiforsystem levelcorrection onfyj Dton* Character Poantona <2-79. SUftbCUftf A019* Sun angles — m« ion elevation angle and aun axtmuin angie maaautaa aeatwea tram trua Norm at tha ttma ol midpoint at mss frame 1* soeofted <0 the naaraai degree- Blank for seconding n«da coverago, Character poaraon 42 givaa ma proeaemng procedure: H*T ■ normal processing "A**■ abnormalprocessing procedure g. Cftaracior position 74 defines m e type of geometric correction 1 0 0 * 0 0 to ma oat* ~ir * unoorraetad *9' • aywan level -Q- ■ geometneeey eor* ractad baaad on geometric CCP-* ■*R“ • gaomartcaitv car* ractad eaaad an rtMova Character positian 44 dabnaa enamar an Earm naoa oran RBV eaabraben ■t- * unaa " T • comi n. Charaatar Posmona 44*100 NABABLANOBATir idanbbaa me agency and 1. Character its Positions 101- B*l 1442*14432-4 Caeh acane has « umowa identification ntimosr, iwnoao format is SwkDOO* HHUMS W W C "£” *> encoded Protect 1 - unoaat 1 2 » Lanout 2 3 ■ undaat 3 □ ODD - Day number rem ove 10 launcn atnma of HM - Hour atom a ofobser­ vation m m * Minute ai ifme of 3 ■ Tana of seconds at lima of obaarution Each image or subacano n toenofiad by a number or letter separated from ma scene 10 by a dean, representing me spectral bond or suoocano iRBV 1 . 2.3;MBS 4, 5,4. 7.4:RBV A, B. C. 0) . _ . •(arm imag* Ewier *t*. or 7 T • RBV nauutieiK caabra* lion imagta. maKsnng lowestto mgnaat aepoeiee level rasoacovefy. a c re . mi NOTtil Q w THBtrrm tB -a* THROOOH-r MmTOPAIUOIIAPHBIII ntii oocumtfT t h a t o>uu>a t w « a w w o t a t w w bcqciu Q CtUMOTM ■MtmON* IMTM, UMMTATMN ILOCIL Figure A-26 (cont'd.). $ Character poamen 45 nw dicstaime aonaor gam op* bont *H* • Mgn gam V* • tow g«n llOQflm a lifm i S S S w RbiiiOA ^ a, Character Pea*ana 42*41 Character! m mta group 111IIIII*a M f ----- h ---a a saaifjissraaa ttMsmaemet- Character pompon 77 I! 34 II 17 7 .1 31 7 .0 35 .5 31 t 37 11 31 31 JO I II .1 31 r 31 .7 31 !* • 19 t .l 11 38 31 1* 7 .C 18 1 35 .9 141 t 3T f It Ji 1* 34 14 34 t.c IT 1 .0 IT ■1 71 5 .0 21 78 79 29 79 7 .1 IS 1 .o 29 .5 78 71 78 21 2 19 .9 s . : 10 49 9 .8 49 1 .0 41 IT 49 1 45 T .l 41 1 .3 <8 !•< 41 I 47 f 40 31 1 t IB 4 .0 la *5 30 30 30 78 71 1 .0 II 11 78 21 79 21 2 T 74 78 .1 28 78 .s 79 79 71 29 28 1C .4 7^ 76 71 25 2 7] 30 li .3 21 21 .1 21 b n o h t v ii a n o n (m c .5 30 T H .1 7 * 5 291 10 IN O n m u in O n 6 * .1 37 I 10 T 71 » .l 11 31 1 .7 11 IH n lN l I N O N D tC 7! 29 T 79 l.C 29 21 l.C 79 l.C 29 T 71 79 29 78 79 a n o n m u in on c n c 4! .5 49 .9 49 l.C 4} .3 4J l.C 41 t . t 17 1 .0 44 r 42 40 1 .0 41 .1 10 a n O N rn iL » n on o n e II T 79 29 1 .0 11 r 21 7*0 78 T 71 78 7< 29 1 75 74 1 .< 7! 14 74 i 1 t lu o m iH i i n ( n cm c .1 Z! .9 79 .1 74 2 .1 21 .2 34 .1 14 21 21 .8 71 m o M n ti in D N CMC i: 1 73 r 77 l.C 71 r 27 l.C 11 7< 71 23 7) 7! tN O M rm v a n o n bad t.c 47 .9 40 .1 ID .0 IT It l.C 31 1 .1 18 l.C 10 1 -5 IN 1 .0 4[ 1 H n l.C !« 1 .1 18 7 .1 SI 2 .0 79 r 79 1 .9 )l 37 z .o 17 1 .7 11 1 .4 14 71 1 .3 l.C 39 1 .7 l.l IS l.C r .3 30 29 1 . 0 fl.C 1. c 1 .1 14 7 .C 19 l* D 29 i.a 31 11 l.l 7« 7 .0 79 7 .1 71 79 71 ?1 3 .4 IS I! 1 .0 71 4 .0 77 4 .0 IQ 30 21 71 71 71 3 .0 IQ IS 79 79 2 .E IT t t 34 7 .0 n i .a 7« iw tf in i m ON 0*0 H 1 14 5 NON* N IL SN ON C *C 21 1! 1C l.C 3C 30 30 1 .9 21 1 .1 1 .2 11 1 4 . C 7 .C 4| 79 11 l.C 31 1 .0 11 1 -9 40 1 .0 47 1 *1 41 38 11 11 79 1 .5 39 *1 II ii 32 JI 17 1 U i.a ■c l.C 42 7 .0 4 | 42 41 11 18 7*0 40 J .C 4 | B N f lH f « C L s n O N CM3 M l* I C fc lJ l 15 T .6 1* 1 .9 1 .1 74 7 .9 S n O n 'A t I 9 N ON cu e 1 .1 71 ti .9 74 34 0 * 4 A ll $N ON e n d 24 29 1 39 1 .7 It it 9 .0 11 I .a 31 4 . 0 2 .c 37 121 2 .9 H ’3 4| 40 IC IB IB IT 1 I .4 38 is l.C 38 4 .2 31 1 .4 18 1 .) IS 18 SI 34 J) 9 .1 J8 r P 34 1 J» Ji II 78 71 t V 32 31 30 1 .0 17 IQ 78 1 71 79 15 74 24 4 | 4 .7 46 ll 49 41 42 41 40 18 )« 21 6 .8 J .j 7T 9 .8 30 9 .1 f 10 T .l 1 21 9 .2 11 a .c 10 8 .0 71 s .o 28 9 .0 .5 li 1*1 P 16 IS is 14 |4 11 11 18 IS 1 .0 » 19 1 .2 SI .8 10 1 -0 30 f JO 1 sc .S 11 9 .1 35 4 .C 19 r 19 34 11 1 <7 34 1 ■! 34 7 .9 40 4 .0 43 .8 is 1 .c 441 T to 7 1 « . c • .0 .1 71 9 .0 .( IS 9 .1 7 .1 71 9 .1 7 .9 )C 9 .1 .4 9 .1 .1 i ■; ii 111 7 -C SI 2 1 1! 2 .9 16 1« 15 ll *■ 70 ,i 28 1 .( 21 3 .0 30 1 .0 31 30 IQ 30 21 4 .0 31 37 79 78 71 71 I 76 )i 13 .5 S3 1 .1 3! 1 31 4 .G 15 .5 3t 3C l.C 30 7 .5 30 28 76 36 16 71 75 J .C 11 t.c 44 31 IT 34 II 77 70 70 70 24 17 14 IS T 15 14 .5 11 1 .0 1 .9 13 1 .2 11 34 4 . a 7 .0 19 16 39 1 .0 IS l.C ij l.C 4] 1 .0 1! 1 C 15 .9 i : 13 1 4] 4? 7>C IJ LA nt C itr SnohM l i SN ON O N ( 4 .0 74 r 77 f 31 •S 72 .2 77 .7 77 73 .9 27 7: 1 77 73 77 77 .1 21 77 S n O w fA lt f N ON ONE r 71 It 1 .1 2! 1 .4 71 ti . ] tO .4 79 1 .3 IS .« IS ■8 li .7 1* 3] 31 .9 11 7 .1 It iN O M tA L t 3 N O N CNC 1 .9 74 1 74 Zi 7 .8 71 11 .1 78 7 .7 79 I .£ 24 .4 78 .1 71 t 78 1 71 t 79 *4 71 IH O tf R lt BN ON o n e 1 .0 34 I 31 ■S 31 J .c 7 -c Si 33 1 .£ |i 1 .c 1! )l 14 .1 31 r 39 17 r 13 1*5 32 S N Q ttm u S N ON ON£ u rn to u jT I .) 21 1 .1 .1 71 ) .< .? 71 4 .7 T 71 l.l s .r r 2< l.l I 1C 3 .1 T 70 1 .7 1*1 31 1 .T J .t 70 4 .4 f 70 4 .7 1 .1 71 4 .4 S N O u tn iL S N O N CN O Jj 17 17 1 ■< 1! 3 .: 3i 1 .0 11 l« JC 19 15 li H 1) >3 1 .a II 79 7S .1 21 1 7C r 76 t 79 1 79 79 r 76 26 79 .5 79 r 79 70 4 .7 *1 70 i .a .1 K 9 .0 1 7C 9 .0 .3 70 4 .7 ,1 70 1 .9 .1 70 4 .1 io 4 .T .1 I* 4 .4 IN 4 .1 1 IS 4 .1 .1 19 4 .1 1* 4 .9 32 3? 1 .0 31 .9 IS .5 29 *> 21 tt+ N t c iit ru n T lttf a? tm « iO u tn O nn n O u C v ito n m « t tiU ** M Tnna t in n t 11 1 • c cM n rm 9 * ,tM lN H f ir tn i* « mm L U O IM O IO H 4 I f n O ftT A O U t * m m e l m to 78 S .« 44 41 21 8 .6 1 71 6 .8 1 1 17 10 T 44 44 1 44 41 4 .0 44 43 40 t n 3 .0 77 72 77 71 21 7 .8 73 t .0 74 7* 77 4 .4 32 ■i 35 33 JI H 10 9 .4 30 T 12 JC 71 25 IS j .; 1! . 1 78 75 24 23 72 1 .5 32 72 1 It I 19 15 IS 31 1 .5 32 12 31 11 ll l.C 32 1 .0 37 79 27 27 13 t* S 11 4 .0 .1 38 4 .1 76 4 .1 79 4 .1 79 i.l 76 4 .1 7 -3 28 4*1 76 4 .7 25 4 .? t 27 4*7 77 1*7 14 3 .0 31 *1 IT IT 16 l.C 15 30 P 1 1 74 1 *0 71 7 .5 31 II 79 79 79 1 .5 78 1 78 71 76 79 3S I) 77 1 3 .1 IV 71 • *T 4 . 9 II 4 .8 71 4 ,8 71 4 .9 70 4 .T 7 .J 7C 4 .5 71 4 .1 70 9 .7 IS 4 .6 IS 4 .1 16 9 .7 i: l.C 11 II 39 18 19 I .c IS 15 31 JC 71 71 28 71 i .i 71 4 .0 11 33 31 3) 31 1 *1 31 1 .3 11 31 30 71 28 38 31 29 1 .1 21 l 76 79 76 75 l.l 78 1 .3 28 71 76 79 71 32 20 is I .S 79 75 IS T 15 t 15 1 .9 2J 1*7 21 IC 18 31 1 31 34 4*0 38 Ji >0 71 71 79 74 27 11 1 35 73 l.i 70 70 11 11 11 16 15 14 1 13 It 19 04 a t C * M U nO v m t « 4*1 21 l.C 11 « 75 31 1 .0 IB in n Q C H iiil 27 4 .1 31 2 .C 11 n lD ; g 13 79 lO u O 75 14 l.C ]l "A ct 74 3* n m io 21 19 3 .0 Si < k t(N N 27 i.a s« su w ro L i » 0 4 C 4 ( .1 t* SN O nT A u I N O N CN O * « n o n 'n f A i T 21 1 .4 35 l.l Si v n n n n fO « l 2 n t h C L L I 'O N 20 3 .a 11 .« 1* 1 30 t m ic m c 19 .6 31 .a 14 S N O N fA U 3 « Om 0 * 0 C H M iH O I * r* A IB IB as lo m m to u s t(U f* D N n IS It 12 13 18 10 ii tM M R il S N ON CNO T o * o < in * fN Q n r < t i i to ik t 2 a n e w all M M WC 5 n 0 m4 * u an on m e I n U n it *0"! ft ft S o? M u tC f c***o 4 3 d i ro t cnt* am n a v e * to n 7 * 1 S N O M n n iL S N O N CMC 7! 5N O W 4A U SN ON C M u rn (Q u it 1 7a 4 .0 IW M N I S N O N ONC t ir .1 ti It l .l tr 3 .1 17 I 33 1 17 f 11 T II 1 II 11 1*0 17 1 .7 7* *9 71 71 1 .0 7* 2*4 10 .1 II 11 *1 SO 71 71 1 It .1 7» J» t 2* 2 .1 IT I . ? 1 1 .4 79 21 21 1 JT 21 79 79 t* i 78 ,7 17 1 .7 71 3 .0 1* 7 .T 11 1 li l* C 7! 7 .0 71 * .o w S n O H T A ti a n on one if 21 s n o w all I N ON UNO i is t 37 * .2 31 4 .1 20 4 .1 1 18 11 5 .4 4 .0 o i sh o w a n on m i r .1 71 29 S N O U fN U I N ON CNO M 32 Ji 1 .o JI f 17 1 .0 11 11 7 -0 IS S N O H f N tL in o n o n o 1*1 .9 21 .7 71 7 .* 13 .1 1) I 73 1 .9 74 cno 13 11 I.N 2T r 71 71 78 t.a I II SI 31 1 31 l .l II I .c IS 71 1 29 79 I 13 .1 78 1 .3 21 1 .9 21 IS 74 .1 .9 11 □ 1 » 71 IS 39 30 11 188 MICHIGAN UPPER ! .! i 1 J '■JsstftiSsT'""!' '7 Slower — iin*ctKTRj j —wein— --ii^ -c k frm A L (Jwer /v CEfjTRAL T' ■* -i T _P i r - ' L - LCWER , : rv LCViER Figure B-l. Location of weather stations 1979). (from NOAA / 189 MICHIGAN E J^ UPPER i Crtater'. Chan?5Z Percentage of yearai during which a 6-inch pr greater; snow depth occurred, i : I LOWER '* . » 50UTH(CENTRALLOAK1 (EH t-m Figure B-2. Percentage of years during which a 6 -inch or greater snow depth occurred (from Strommen, 1968). 190 MICHIGAN ’^f'cfNMAQLCWEB iTHEftST SOUTH ""SO U T H CENTRAL LQWt Figure B-3. Average number of days per season with accumulated snow depth on the ground of 6 inches or more (from Strommen, 1974). APPENDIX C DIGITAL CLASSIFIC ATION OF LANDSAT MULT ISP ECTRAL SCANNER DATA A. Introduction to Image Classification A f u n d a m e n t a l , and often primary, goal of remote % sensing analysis is the classification of an image scene). Classification, into discrete, or the partitioning of the image pre-determined classes, can be accomplished either manuall y (photo or image interpretation) aid of computers (or (digital processing). require the use of decision rules. or with the Both approaches These rules (or keys) are ge ner ally derived from an analysis of areas considered representative of the various classes. utilize color, image elements texture, (e.g. pattern, size, site, shape, shadow, and association) characterize a particular class, rely on numerical Photo interpreters tone or to whereas digital techniques parameters. Classification by both techniques will be illustrated for a Landsat MSS sub-scene acquired on February 26, over Wexford County, Michigan (Figure C-l). 1979 A fairly direct visual classification technique would characterize each class by tone (white = snow, gra y = hardwoods, p i n e s ) , and then partition the image tonal comparisons. The decision rule and black = into classes based upon is simple; area to the class with the same, or most similar, 191 assign each gray 192 Figure C-l. Landsat image (E - 3 0 3 5 Q - 1 5 4 7 1 ) , black-andwhite band 5 (S-snow, H-hard wo ods (leafless), P-pines). 193 tone. The numeric equivalent of this technique would utilize digital values {obtained from a densitometer or brightness values tones from the Landsat CTT) {Figure C-2). mathematically; in place of gray The decision rule may now be stated assign the unknown area x to the class to which the distance is nu mer ical ly minimum (H). Distances are simply the absolute value of the unknown area minus the class to which it is being compared d = | x - H (Figure C-2): | The use of tone in the above analysis was chosen since it illustrated the concept of spectral pattern recognition. Of the three major image characteristics and temporal a t t r i b u t e s ) , spectral (spectral, spatial, patterns are the most commonly utilized feature in digital classification. B. Spectral Pattern Recognition 1. A Geometric Interpretation When more than one band of Landsat MSS data are to be analyzed, individual bands may be represented by color images and superimposed to create color composite This technique is limited to three bands, composite, for any single and is also not capable of rendering range of tonal images. the full information available from the MSS data. Digital classific ation techniques utilize the full range of spectral data (brightness values) bands simultaneously. available from multiple 194 SNO W X 1 J WHITE | H AR D W O O D S 1 PINES 1 GRAY BLACK 1 t I I I l 1 l 120 110 100 80 60 ■■ 40 20 0 * Figure C-2. * Correspondence between Landsat MSS image gray scale and digital brightness values from CTT. 195 Although the computer operates strictly in a numerical mode, graphical techniques will be used to introduce geometric concepts in image classification (this technique is adopted from one presented by Lillesand and Kiefer (1979)). The portrayal of spectral responses/ spectral patterns, is limited to two dimensions, computer implementation of these techniques and can be applied is mathematical brightness values from the CCT corresponding to the sub-scene Figure C-l. microcomputer.) in the section on the ERDAS Individual areas, called were identified from ancillary data sources photography) in (Actual computer procedures for data manipulation are covered types. whereas to almost any number of bands. All subsequent data represent digital extracted and thus as being These sites, "training sites," (maps and aerial "representative" of the various cover therefore, represent a sample of pixels which will be utilized to characterize individual cover types. (numerically) the Individual pixel values have been plotted onto a two-dimensional graph and are identified by letters indicating to which category the value pertains (Figure C-3). Band 6 brightness values are plotted on the x axis with its corresponding value in band 5 on the y axis. Visually, values, each class displays a variable range of possible although the spectral responses tend to form discernable patterns. These three training sites will be utilized to illustrate several classification strategies. 196 A Value 40- Band 5 Brightness 50- 30 - 20 H HH H Hardwoods (leafless) hm hh ** H H H - J J J J J Jack p i n e ^ ^ J R R R RR R R R R R RR R RR R RRR RRR H H , ^ R e d pine R RR R 1--------- 1--------- 1--------- 1--------- r ~ * ~ 10 20 30 40 Band 6 Brightness Value Figure C-3. Two-dimensional values. 50 plot of training site pixel 197 An extension of the tonal classification illustrated in Figures C-l and C-2/ technique/ utilizes two, or more/ brightness values (the numerical equivalent of gr ay tones). Each class is charac ter ized by a set of mean values obtained from the pixels located within the training areas in each spectral band. the one band The decision rule is a direct extension of (gray tone) technique; assign an unknown pixel to the class with the same/ or most similar/ brightness values (gray tones). This strategy/ distance-to -me ans classifier/ known as the minimum is illustrated in Figure C - 4 . Class means were determined and those for band 5 and band 6 are shown as +s on Figure C-4. pixel/ To classify an unknown such as points 1 / 2, and 3/ the distance between the unknown pixel and the mean value for each class is determined. connecting For point 1/ these distances are shown by lines the point to each class mean position. decision rule would be: "closest" assign the unknown (i.e. minimum distance) class/ The pixel to the Point 1 would therefore be classified as belonging to the hardwood class. The mat hematics for expressing the distance measure nu merically is illustrated in Figure C-5. and y (BV5) axes are perpendicular/ for right triangles can be used. hypotenuse is equal Since the x (B V 6 ) the Pythagorean theorem Since the square of the to the sum of the squares of the sides: 198 50- Band 5 Brightness Value H H HH H 40 - HiH H R R R R R R 20 R R (Fit - RRR RRR RR 20 30 40 Band 6 Brightness Value 10 Figure C-4. 50 Geometric interpretation of the minimum distance-to-mean s classifier. 199 Mean value (or y a s p e c i f i c c la ss (BVe.B^B) Unknown pixel, (BV0.BV6) / x bv Figure C-5. e Geometric derivation of distance. 200 d = V ( x2“xl)2+(Y 2 “y i )2 For n dimensions/ or bands# the generalization <3= V (x 2 “ x i )2 + (Y 2 - y i )2 + ... is: (n 2 - n i )2 or# expressed in terms of BVs: d = Therefore# n E i= l __ {BVi - B V i )2 the formulae for calculating the distance between an unknown pixel and the mean value for a specific class# using 4 band MSS data# would be: d = V(BV4-BV4)2 = (BV5-BV5)2 + (BV6-BV6)2 + (BV7 - B V 7 )2 Note also that the above a = V( formula for one dimension# x 2 - x i )2 # is equivalent developed to the distance measure for the one band tonal classification. the calculation of distance as presented above# Euclidean distance# known as may intuitively be the "best" measure# other distance measures have been applied data analysis Although (Swain and Davis# in remote 1978). While the minimum distance-to-means classifier relat ive ly straight-forward, sensing is its reliance on mean values alone to characterize classes ignores the variance exhibited within the classes. For example, point 2 (Figure C-4) would be classified as belonging to jack pine# mean# when, in fact# it "appears" the closest class to belong to the red pine 201 class. The classification strategy is thus insensitive to differential variance exhibited by classes. Statistically, the simplest measure of va riabilit y for a set of data values is the range. brightness values (i.e. range) The minimum and maximum are obtained from the training set and are used to bound a category. rectangular decision region, the class range (Figure C- 6 ). straight-forward; for each class, Thus, a is defined by The decision rule is assign an unknown pixel to the class decision region in which it occurs. measure of variability, By introducing a this strategy would now assign point 2 (Figure C- 6 ) to the red pine class. Note that point 1 is also classified as red pine and that point 3 lies in two decision regions. Overlapping decision regions commonly occur whenever classes display correlation between bands. Both hardwoods and red pine are highly correlated in band 5 and band 6 , thus producing a positively slanted series of pixel observations. It can be seen that highly correlated categories are ill-defined by rectangular decision regions. The more general situation for this classifier occurs whenever opposite sides of the class bo undaries are parallel to each other, but not necess aril y to the coordinate axis, producing parallelograms or, (Figure C-7) (S c h o w e n g e r d t , 1983) for multi-dimensional data sets, parallelepipeds. 202 Value 40 - Band 5 Brightness 50- 30 20 - 10 - H H H H H H H H H H H HH H H -3 R 1• H RR RR RRR RRRR RRRR RRR RRR i j j 2• R RR J j j R j ~i--------i-------- 1-------- r* 10 20 30 40 Band 6 Brightness Value Figure C - 6 . 50 Geometric interpretation of the parallelepiped classifier, with rectangular decision regions. 203 Value 40 - Band 5 Brightness 50 - 30 - HHH H H / H H HH RR RR / R R R /R R R R 20 - J J J 10 RRRR RRR RRR R R - “ i-------- 1-------- 1-------- r~ 10 20 30 40 Band 6 Brightness Value Figure C-7. 50 Geometric interpretation of the parallelepiped classifier with parallelogram decision regions. 204 2. The Statistical Approach Statistical decision geometric theory/ as opposed to the techniques presented above/ quan tit atively account for both variance and correlation in the data set. parametric techniques/ With the distribution of the various categories must be specified/ with the most common assumption being that the data distributed (Moik/ Swain and Davis/ 1980/ is no rmally (Gaussion) Lillesand and Kiefer/ 1978). The probability de ns it y functions are used to compute the statistical pr obability that an unknown pixel belon gs to a particular category. decision rule is to assign the unknown pixel with the highest probability. closer to the category to the more variable red pine The estimated pro bability density function = (2 for the (i.e. considering only one band of data as in Figure C- 8 ) is: p(x|ci) is for the hardwood class but has a higher probability of belonging univariate case the function as illustrated Note that point 3/ from Figure C-4/ to the mean class. The For the univariate case/ maximum likelihood classifier will in Figure C- 8 . 1979/ and 1 ) 1/2 Si exp - 1 (x-mi ) 2 2 Si2 where: p (x|ci) = probability that x is from classi m£ = the mean BV for classi Si = the variance for classi 205 Hardwood p<3 I R P ) - p(3 I H ) - ‘ 22 Mean (Red pine) Figure C - 8 . 31 Point 3 38 Me a n (Hardwood) B a n d 5 BV Classification based upon probability density functions. 206 and the decision rule is: classy decide that x is a member of if and only if: p(xlcji) >_ p (x|cj) for all j. Note that each cat egory is characterized ent irely by its mean(s) and variance {-covariance). Wheneve r two bands of data are analyzed, probab ili ty den sity function the bivariate is utilized to compute probabilities: 1 p{ *1/X2Ic i ) = T 7 2------ (s ills i 22 “ s i 2 1 2 ) 2 (xi - mix)2 2sii2(xi-mi!)(x2-mi2) (x2-mi2)2 + s ill s ills i 22 x exp - - s i 22 silisi 22 where: mij = the mean BV in band j {for classi) sijj = the variance in band j (for classi) 3 ijk = the covariance between bands j and k { for classi) The pr oba bil ity den s it y function for two bands may be thought of as defining a series of points of equal probab ili ty about the mean (Figure C-9). The center of a class is deter min ed by the mean and the shape by the covariance matrix, points of equal probab ility are therefore 207 50- 40 - H H, H H S 30- 20 - L i n e s of e q u a l probability R R Jt R R RRR ■o 10- 20 30 40 10 Band 6 Brightness Value Figure C-9. 50 Points of equal probability as defined by the maximum likelihood classifier. 208 elliptical (hyperellipsoids for more than two bands). The maximum likelihood decision rule would classify points 2 and 3 as red pine and point 1 as hardwood (Figure C-9). Probability density functions for more than two bands are presented in the section on matrix notation. The formal derivation of the maximum likelihood decision rule is based upon the Bayesian principal to minimize the average loss over the entire classification process (Swain and Davis, 1978). The Bayesian technique is theoretically an optimum classifier which applies two additional weighing factors to a probability; an a priori probab ili ty of occurrence and a loss function. often the case, assumed equal; cation If, as is the a priori probabilities are unknown and and the loss due to an incorrect cl ass ifi­ is simply defined as being inversely proportional the a prior probability, to the Bayes optimal strategy ge neralizes to the maximum likelihood classifier previously presented. Matrix algebra provides a means for condensing large mathematical manipulations symbols (Searle, manipulations multivariate Therefore, 1966). into a much smaller set of Vectors, (matrix algebra) matrices, and their are commonly used in statistical analysis (Morrison, 1976). many of the image classification techniques will be presented in matrix notation. 209 A Landsat scene is recorded on a computer compatible tape as a matrix, vectors, L, composed of four-dimensional spectral Pij, which represents brightness values of individuals pixels in four MSS bands: 1,334-8 1,1 L - P 2 9 8 3 , l ........................ p 2983 ,3584 where L = Landsat MSS scene; and BVltij BV5ij BV6.. B V 7ij where P^j = four-dimensional vector corresponding to a single pixel and defined by the digital counts {or brightness values) in the four MSS bands. Pixels located within a training site, xc , are used to estimate a mean spectral vector for a specific category: 210 X1 where x2 Xi = |^BV4i BV5 i BV6 i BV7iJ xc - and n = sample size The mean spectral vector, arithmetic means, bands, Xc , of class C, is defined as the BVC , of the brightness values, in four from the pixels within the training site: BV 4 ( BV5( BV6C BV7, The computation of the mean spectral vector is: Xc = n (Xc ’l) The varia nc e-c ovar ian ce matrix, Sc , for a training site class is defined as follows: v a r (B V 4 ) S- = c o v ( B V 4 ,B V 5 ) c o v {B V 4 , B V 6 ) c O v ( B V 4 ,B V 7 ) v a r (B V 5 ) c o v (B V 5 ,BV 6 ) c o v (B V 5 ,B V 7 ) v a r (BV 6 ) COV(BV6 ,BV7) v a r (B V 7 ) 211 The computation of the va ria nc e-c ovarianc e matrix Sc = is: X' C XC - - (x'cl)tl'Xc) n n -1 The probability density function/ employed in the maximum likelihood classifier, for a particular class is: p(Pij|C) |^exp-l/2(Pij-Xc )-Sc- l ( P i j - X c )J = (2 where: it )n /2 |s p(P£j|c) c 1 1/2 = pro bability of pixel Pij belonging class C n = dimension of vector, Landsat MSS data) Pw to (4 in the case of Xc = mean spectral vector for class C Sc = va riance-covariance matrix As expansion of the computational will illustrate the methodology. for class C forms for Xc and Sc Assume that a training site consisted of only two sample pixels: x„ c = BV4 BV5 BV 6 BV7 ~X 11 x 12 x13 X 14_ Pixel 1 X 21 X 22 X23 X24 Pixel 2 n = sample size n* 4 212 The computation of the mean spectral vector would be as follows: xc = i ex'c 1 ) n X11 x 21 x 12 X22 x 13 x 23 X14 x 24 X11 + X21 X12 + X22 X13 + x 23 + x “14 24 -* 1 L J n* ! 4* n £xil £XU 1 n Ex i4 4-1 4-1 BV4 Exi l ^ n £xiJ c BVSc Exi 2 / n £Xi 3/ n BV6C Zxi 4 / n ^ u BV7c J 4-1 4-1 213 The computation of the variance-covariance matrix would be as follows: Sc = n - 1 [ Xc Xc ' n C X ^ l ) ( l ’ x c )J Zx i l ‘ ^ x i l ^ 2 _____________ n Cxilxi2 - n-1 CExi2D n-1 4•4 var(xi l ) covCxiltxi2) var(xi2J S_ = covtxil* x i3 ) cov(xiltxi4) c°v(xi2 ,xu ) cov(Xi2 ,xi4) var(xi3 ) cov{Xi3,xi4) var(xi4 ) 4-4 var(BV4) S_ = cov(BV4,BV5) c o v ( B V 4 , BV6) cov(BV4,BV7) v a r ( B VS ) c o v { B V S , BV6) c o v [ B VS , B V7 ) var(BV6) cov(BV6,BV7) var(BV7) 4-4 T *fr t-TX3 £T XI I^X ZT X I I T XI **x « x £ IX £ZX £ I X zzx ZIX zzx Z IX TZX n IZ X n « x « x £ZX I x x’x x t-.fr t -T X 3 £ TX I ^x^x2 D D X X ^X^X3 ^x^xi tTxTTX2 £TXI T I X3 IT XI » fr.fr H x+H x z + z fr Zx £ Z x + n x £ I x £ZX+£IX tf Z x Z Z x + tr I x Z I x £ Z x ZZx + £ I X Z I x z z zx+ z ZZXIZX+ZIX IIX £ZxIZx+£IxIIx « x ™ x + ” x™x Vx ZZVjtZIv I^x+IJx Z Z — U ■ fr frZx « x IZ X £ZX ^ x ™ x zzx Z IX IZ X " x (j*u « x ” x £ZX zzx £ IX Z IX *\z - o o X X I 215 216 Exil " CExii:)2 _________ n n -1 z x ilx i2 * C E * i i ) C E x i 2 ) n- 1 4*4 varfx^) COVtXij.Xij) c° v ( x i l f x 13) covCxu , x i 4 ) var(xi2) covCXi2,xi3) cov(xi2 ,xi4) var(xi3 ) coV Cxi3 ,Xi4J var(xi4) 4-4 var(BV4) cov(BV4,BVS) cov(BV4,BV6) cov(BV4, BV7) varfBVS) cov(BV5, BV6) cov(BV5,BV7) var(BV6) cov(BV6, BV7) var(BV7) 4*4 217 C. Unsupervised Clustering The clustering algorithm/ micro-computer/ set. involves two passes through the entire data During the first pa s s t clusters are created by an iterative process. pass/ as implemented on the ERDAS The final cluster means/ are then utilized from the first to assign each pixel in the data set to a cluster based on quadratic distance. Figure C-10 is a diagrametric representation of the method utilized to create clusters during the first pass. Pixel values are read/ one at a time/ and assigned to an existing cluster or used to create a new cluster based on the following decision rules: if r < rmax if r > rmax / p isassigned to if Cn < Cfj/ a new cluster a current cluster using rm in is formed if Cn = C n c small < Cm in/ c s is deleted and the new cluster replaces if c small cmin* at merger M/distance D it the new pixel P is discarded between all cluster means are computed if D < Dm in / the two clusters are merged systematically 218 U N S U P E R V IS E D r = C LU ST ER IN G distance between pixel P and a cluster mean rmax = a threshold distance rmin = minimum distance between a pixel P and the cluster means number of clusters CN maximum number of clusters c small = cluster with smallest population cmin “ a threshold population number M number of pixels D = to be considered before merger distance between cluster means Dmin = minimum distance between cluster R maximum allowable cluster radius = Note: CN' M ' Dmin' an(3 R are user specified values. Figure C-10. Unsupervised clustering. 219 After the entire data set has been processed, are frozen. cluster means During the second pass pixels are assigned to a cluster based on quadratic distance. If r > R, the pixel is unclassi f i e d . The following example utilizes data from the Wexford County test site (Table C-l) to illustrate the assignment of individual pixels to specific clusters. Figure C-ll is a plot of mean brightness values for the 27 clusters, table C-2 lists the mean brightness values, by band, each cluster. between To classify pixel A, for the quadratic distance it and each cluster mean would be calculated as: d = X / ( w 2 - w x )2 + (X2 - X 1 ) 2 + ( Y 2 - y i ) 2 where: while d = quadratic distance w,x,y,z, = BVs 2 = unknown in four bands pixel 1 = cluster mean + ( z 2“ Z l ) 2 220 Table C-l. Digital values* for bands 4,5*6,7* and 8. BOUNDARY SNOW JACK PINE Pixels Band Row 4 5 6 7 8 4 5 6 7 8 4 5 6 7 8 4 5 6 7 8 80 80 80 80 80 113 126 125 90 1 22 88 12 81 81 81 81 81 82 82 82 82 82 84 84 84 84 84 114 127 111 111 127 18 91 18 127 119 90 14 112 112 110 110 127 125 89 127 123 89 19 127 12 127 123 89 13 111 111 112 112 127 123 127 124 89 13 127 124 127 125 87 14 86 12 116 127 124 88 16 Column 123 122 88 13 122 87 14 116 127 126 A \ 105 127 123 15 74 87 90 65 15 26 27 38 31 15 109 127 124 90 15 89 104 99 73 15 38 42 44 33 15 17 115 127 126 91 15 102 122 110 81 15 47 57 55 41 15 14 14 18 18 15 118 127 125 91 16 72 85 78 64 16 19 19 27 25 16 128 129 130 131 88 116 127 126 90 16 116 127 126 90 14 15 127 127 91 15 124 125 126 127 88 B I 15 16 9 • 11 21 19 120 16 13 16 12 18 20 21 20 21 15 15 15 15 15 16 12 12 20 18 15 17 13 18 20 15 29 15 16 12 20 22 15 12 9 21 18 16 22 22 21 21 15 15 19 16 25 23 15 19 17 23 15 19 16 24 24 16 19 17 25 26 16 22 132 133 221 1 2 7 120 - — Non-Foraat - - Rad •— Jack —• R a d , Ja ok and Mtxad P ina* M o n -F o ra a t with more than 26% F o r a * t •••* 110- Pina Pina 100- 22 Digital Count * ( Br i g h t ! ) * * * Valua*) 9 0 -12 80 - 7 0 - -8 60 23 5 0 - 40 27 21 20 28 13 -28 -2 4 3 0 - B 20- 10- "T 4 t 6 T 6 ~r 7 S p a o lra l Band Figure C-ll. Plot of mean brightness values for 27 clusters, 222 Table C-2. Mean brightness values# by band# for 27 clusters. Cluster #3 1.019 3.484 5.307 46.17 70.37 87.81 96.11 52.78 84.68 96.31 56.60 82.90 72.88 48.02 64.95 #1 % 4 5 6 7 % 4 5 6 7 % 4 5 6 7 #2 #4 #5 1. 374 2.442 88.28 101.26 106.14 110.88 99.56 110.31 81.68 74.85 #7 3.597 6.706 36.09 42.79 47.78 38.22 49.68 48.54 44.82 41.43 #6 .308 77. 50 82. 20 87.68 6 8 . 36 #9 2.444 92.56 110.48 106.71 75. 53 #8 #13 2.537 28.57 29.55 33.44 30.26 #14 2.531 74.31 88.57 85.09 66.24 #15 4.621 35.38 37.96 40. 58 34.98 #15 8.127 80. 18 95.63 92. 16 70.74 #17 2.064 122.43 127.00 126.95 97.21 #18 1.176 15.82 12. 14 21.08 2 2 . 10 #21 #22 5.427 24.68 23.35 36.17 36.00 8.354 114.73 126.97 126.33 92. 27 #23 3.615 65.05 77.00 74.49 58.49 #24 3.479 15.76 12.79 24. 50 26.58 #25 3.648 15.90 12.79 29. 14 32.76 #26 2.691 31.00 30.00 3 5.00 31.00 #27 4.054 37.79 41.04 43.08 37.58 #10 #11 #12 6.585 56.45 59.90 61.19 48.64 4.565 49.98 57.52 58.13 47. 27 3.707 106.24 116.51 115.91 85.68 #19 6.801 61.59 72.91 72.88 58.13 #20 4.336 19.18 17.68 32.59 34.99 223 the following values would be obtained: cluster no. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 d 46.82 122.38 71.21 21.79 37.91 133.01 144.56 66.31 27.98 110.58 117.14 12.93 163.38 65.80 149.60 53.38 20.21 189.41 90.63 174.99 165.64 11. 13 85. 56 185.70 181.76 160.87 224 Based on these measures of distance/ classified as belonging pixel A would be to cluster number 22. that pixel A is least like Note also (maximum distance) cluster number 18/ which is most like pixel B/ jack pine. Upon completion of assigning each pixel to a cluster the analyst must assign appropriate category labels to each cluster or group of clusters. This is usually accomplished by reference to supplemental data such as existing maps or aerial photography. D. Landsat Classification on the ERDAS Microcomputer The Earth Resource Data Analysis System (ERDAS) features both a grid-based geographic information software package and a Landsat digital analysis software package. The system is a stand-alone computer system based around a Z-80 Central Processing Unit/ FORTRAN/ supporting ASSEMBLER/ and PASCAL programming languages. includes dual double density/ double sided drives/ a joystick cursor control/ color monitor/ and tape drive. BASIC/ Other hardware floppy disk CRT/ matrix printer/ The Landsat software RGB package is a modular interactive system permitting color display of a 240 by 256 pixel/ three band/ false color image of a user selected subscene. The Landsat software includes supervised training site selection/ visual enhancement/ correction/ haze maximum likelihood classification/ unsupervised classification. and Output is in the form of a color disp lay and computer line printer maps. 225 Landsat classification is accomplished by a series of modular programs (Figure C-12, actual program names are shown in all capitals within the terminal/interrupt symbols) executed in various sequences depending upon the particular analysis desired. All processing is accomplished on a Landsat file which is a subset of a standard Landsat MSS scene. The most common method of creating a Landsat file is to extract a portion of the scene from a CCT using program LOADBIL. This program will load any rectangular subset of the scene with appropriate line and column numbers determined from a transparent grid aligned over a corresponding Landsat image (Figure C-13). Also, of bands from the tape, up to a maximum of eight, selectively loaded and in any order. with CCTs produced with band formats. pixel (BSQ) can be is utilized interrelated-by-1ine CCTs with a band-sequential LOADBSQ program, and older, LOADBIL any number (BIL) format use the " X- form at ," band-interleaved-by- ( B I P ) , use the LOADX program. In addition, a Landsat file can be created from a floppy disk which has been purchased for a specific 7-1/2-minute quadrangle ground area. All subsequent processing of a Landsat file require that statistical information for the file be contained trailer file which is created by the BUILDH program. the execution of BUILDH, listed on the printer in a During statistics and histograms may be (Figure C-14). 226 ERDAS Landsat Classification LO A D B IL BUILDH Statistlos Histogram CLUSTR REA D FCC Image max like FIELD :---------- it-jk. Classified Image QIS File Signature Contlgency Table File ( d isp l a y ) ( E R D A s) Hard Copy Options Figure C-12. Flow diagram illustrating the major components for Landsat classification. 227 0093 0093 0032 DOLE t Ut l CD 2 g 0002 006L 0061 d g 009^ 0091 OOn OOCl1 cost 0011 0001 00 6 < 009' OOi* 009 009 00>ooc< 003' OOi' £ s s § s § * s s i £ E S j It In II tt !H fSi Figure C-13. 3? s i 5 § £r- 1 ! i I E S CC CM LU % K z Tr ansparent grid for determining line and column numbers from a Landsat image. 228 Histogram listing for file T1 This image has 205 ROWS/ 190 COLUMNS This is a LANDSAT data file There are 4 bands in this data set This is for BAND no. 7 Minimum data value Maximum data value Mean value Standard Deviation Median Mode NUMBER DATA OF VALUE POINTS 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 2. 1. 2. 2. 2. 5. 3. 5. 16. 25. 23. 36. 68. 95. 86. 136. 146. 207. 322. 395. 461. 570. 754. 949. 912. 1074. 983. 1006. 1010. 1022. 928. 809. 125 54.104 20.66512 49.000 33.000 % .01 .00 .01 .01 .01 .01 .01 .01 .03 .06 .06 .09 .17 .24 .22 .35 .37 .53 .83 1.01 1.18 1.46 1.94 2.44 2.34 2.76 2.25 2.58 2.59 2.62 2.38 2.08 Figure C-14. 8 = = = = = I I I I I I I I I I I IX IXX IXXX IXXX IXXXXX IXXXXX IXXXXXXX IXXXXXXXXXXX IXXXXXXXXXXXXXX IXXX XXXXXXXXXXXXXX IXX XXX XXXXXXXXXXXXXXXX IXXXXX XXX XXXXXX XXX XXXXXX XXX XX IX XXXX XX XX XXX XXX XXXXXXXX XXXXXXXXXX X IX XXXXXX XXX XXXXXXXX XXX XX XXX XX XXX X I XX XX XX XXXX XX XX XXXX XXX XX XXX XX XXX XXXXXXXX X I XX XX XX XXXX XXX XXXXXXXX XX XXXXXXXX XXX XX I XX XX XX XXXXXXXX XXX XXXXXX XXX XXX XXXXXX XX IX XX XXXXX XX XX XXXX XXX XXXX XXX XXX XXXXXXXX IX XX XX XX XXXXX XX XX XXX XXX XXXXXXX XXX XXXXXX IXX XXX XXXXXX XXX XXXXXXXX XX XXXXXXXX XX IX XXXXXXXX XXX XXXXXXXX XXX XXX XXXX Fig ure Portion of the statistics and histogram listings available from program BUILDH. 229 Both supervised classification techniques require that the user provide training data/ a signature file, (MAXCLAS) in the form of to the system prior to classification. This is accomplished by interacting with the image display to locate areas in the scene which can be cover type. Program READ is utilized bands from a Landsat file. a 240 by 256 pixel image of red, assigned to display up to three The program will display, up to set, each selected band as a gray-scale green, color composite identified by or blue, (FCC) superimposed to create a false image on the screen. Any band may be to any color {red, green, or blue) tion and reduction of the displayed with mag nifica­ image possible. The signature file is created by the interaction of program FIELD with the FCC image and the associated Landsat file (Figure C-12). A movable cursor is utilized to draw a polygon on the screen, by means of the joystick, "bounds" an area of known cover type, which usually determined from existing maps or aerial photography (Figure C-15). Then, based upon an analysis of a histogram (Figure C-16) and summary statistics (Table C - 3 ), the user determines if the site ap pr oxi mately represents the desired cover type. This procedure is continued until several hopefully containing training sites, the full range of values, for each cat egory in the data set. are obtained A variety of signature manipulations are available for creating a finalized 230 Outlining an area of known cover type Figure C-16. Histogram of training site data. 231 Table C-3. Data Value (BV) 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 ERDAS training sample statistics listing, Landsat filename: Tl. HISTOGRAM Listing. Total number of points in this sample - 24 Number of counts per data value by band. ------- ----------------- ----- -----------Band 5 Band 6 Band Band 4 0 0 0 1 3 5 11 3 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 4 15 2 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 3 8 9 3 0 1 0 0 0 0 0. 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 '0 1 2 5 7 6 1 1 1 232 signature file for input to MAXCLAS is utilized to examine, signature file i.e. print, (Table C-4). (Figure C-17). GETSIG the contents of a ADDSIG may be used to synthesize a new class or signature from two or more existing ones. SIGNAM is used to create a list of signature names contained in a signature MAXCLAS. Otherwise, file and is required for signature names may be created using APDSIG which will append signatures from one file onto the end of another. As a further check on the homoge neit y of training samples, program CMATRIX may be utilized to generate a co ntingency table wherein the sample areas are classified using the specified signature file With the previously compiled (signature file), Landsat file. (Table C-5). training site data MAXCLAS may be used to cla ssify the entire The minimum dis tan ce-to-means option calculates the spectral distance between each pixel and the mean value for each training site. The pixel is assigned to the class for which the distance is smallest. likelihood option uses the mean The maximum spectral vector and variance-covariance matrix determined for each training site to determine the probab il ity of occurrence of each pixel into each class. The pixel is assigned to the class for which the pro bability of occurrence is highest. either classifier will be a classified geographic information system (GIS) Output from image stored as a file. This file may be 233 FIELD print contents Signature File signature 1 name s ig n a t u r e 2 name s i g n a t u r e 1 ■+ signature 3 = new signature APDSIG enter names Signature Name File name 1 name 5 New Signature File signature 1 name signature 5 name Figure C-17. Flow diagram of signature manipulation programs. 234 Table C-4. Signature list for file: Tl. The signature name is PJ Polygon coordinates are : XE 13 = 131.000, YE 13 = 76.000 XE 23 = 131.000, YE 23 = 84,000 XC 33 = 134.000, YE 33 = 84.000 XE 43 = 134.000, YE 43 = 76.000 XC 53 = 134,000, YE 53 = 76,000 There are 36 pixe ls within this polygon. Band 4 5 6 7 Minimum 13 9 17 IS Mean 16.19 13.00 20.36 21.25 Standard 1.56 2.07 2.21 1.G9 Maximum 19 19 26 25 Covariance Matrix 4 5 6 7 2.43 1.97 1,96 1.90 1,97 4,28 2.94 2.00 1,96 2.94 4,90 1.53 1.90 2.00 1.SB 2.85 The signature name is PR Polygon coordinates are : XC 13 = 135.000, YE 13 = S3.000 XC 23 = 135.000, YE 23 = 93.000 XE 33 = 137,000, YE 33 = 93.000 XE 43 = 137.000, YE 43 = 38.000 XE 53 = 137,000, YE 53 = 88.000 There are 18 pixets within this polygon* Band 4 5 6 7 Minimum 11 11 28 31 Mean 19.94 18.22 33,72 35.17 Standard 4.49 4,29 3.60 2.50 Maximum 28 26 39 40 Covariance Matrix 4 5 6 7 20,16 18.12 14.54 8.45 13.12 18,39 13,62 7.91 14.54 13.62 12.98 5.93 8.45 7.91 5.93 6.25 The signature name is HWD Polygon coordinates are : XE 13 = 140.000, YE 13 = 76.000 XC 23 = 140.000, YE 23 = 81.000 XC 33 = 145.000, YE 33 = 81.000 XC 43 = 145.000, YE 43 = 76.000 XE 53 = 145.000, YE 53 = 76.000 There are 36 pixe ls within this polygon, Band 4 5 6 7 Minimum 33 32 35 29 Mean 40.75 45.22 47.14 40.61 Standard 4,72 6,16 6.19 5.52 Maximum 50 56 58 49 Covariance Matrix 4 5 6 7 22.24 27,19 26,87 24.46 27.19 38.00 36.41 31.78 26.87 36.41 38.34 33.0 2 24.46 31.78 33.02 30.51 235 Table C-5. Landsat class! ication contingency table. Signature File: Landsat File: Class Total Name Points PJ1 PR1 HWD1 SN0W1 [39] [24] [42] [30] PJ1 39 1 0 0 Tl Tl PRl 1 0 0 .0 % 0 4.2 23 0.0 0.0 2 0 0 ..0 % 95..8 4..8 0. .0 HWDl 0 0 0 .0 % 0.0 40 95.2 0 0.0 SNOWl 0 .0 % 0.0 0.0 30 1 0 0 . 0 0 0 0 236 displayed on the color mon itor {Figure C-18) or printed with one of several symbol/overprint options. Unsupervised clustering proceeds directly from a Landsat file to a classified output/ GIS, image (Figure C-12). file will consist of individual pixels assigned to the various clusters which were established C-19). The (Figure By using the DISPLAY program with color assignments (program PALETTE), ca teg ory labels the analyst must assign appropriate to each cluster or group of clusters. 237 Classified image from the MAXCLAS program Figure C-19. Individual pixel assignment to classes (27) from an unsupervised clustering. 238 APPENDIX D Computer program to calculate a kappa statistic and variance from an nxn co nti nge ncy table. 238 239 program k a p p a (i n p u t / o u t p u t /p r i n t e r ); Program to calculate a kappa statistic (K hat) and its variance (Sigma hat) from an nxn co nti ngency table. Written in PASCAL for an IBM-PC. var printer :text; table sar ra y[ 1 .. 1 0 ,1 .. 1 0 ]of real; rowtot/coltot : ar ra y[l..1 0 ]of real; size/i/j/k sinteger; n ,s i g m a h a t /thetal ireal; t h e t a 2 /t h e t a 3 ,theta4/khat :real; t e m p i /temp2/temp3 :real; ans/printan :char; begin writeln('Do you want the results written to printer? (Y / N ) ’); readlnfprintan); if (printan = 'y') or (printan = 'Y') then begin a s s i g n (p r i n t e r /' 1 s t : 1); rewrite(printer); end; writeln ; w r i t e l n ( 1Please enter the size of the matrix. (1..10)'); readln(size); (******** r ead in the matrix ********) for i := 1 to size do repeat begin writeln; writel n(' For row '/ i t l / 1 of the table enter the d a t a .'); for j := 1 to size do begin w r i t e l n ('Enter column '/j rl /' .’)? readln(table[i,j]); end; writeln; w r it e ln f 'I s th is data correct? w riteln; for j := 1 to size do (Y/N)')* w r i t e ( t a b l e t i /j] : 7 : 3 , 1 '); writeln; r e a d l n ( a n s ); end; until (ans = 'y') or (ans = 'Y'); if (printan = 'y 1) or (printan = 'Y') then for i := 1 to size do begin for j := 1 to size do 240 w r i t e (p r i n t e r ,t a b l e [i ,j ] :7 :3 , ' ' ); writelnf p r i n t e r ) ; end; (*calculate totals for the rows, columns, and entire table*) n := 0 .0 ; for i := 1 to size do begin coltotCi] := 0 .0 ; rowtot[i] := 0 .0 ; end; for i := 1 to size do for j := 1 to size do begin n := n + t a b l e [ i ,j ] ; rowtot[i] := rowtot[i] + table[i,j]; coltotTj] := coltot[j] + table[i,j]; end; writeln; writelnf'n = ',n:6:5); for i := 1 to size do wr ite lnf 'R ow ' ,i: l , 1 total = ',r o w t o t [ i ] :6 :5, ' Column ',i:l,' total = ',col t o t [i ]: 6 :5); (******* calculate the thetas *********) thetal := 0 .0 ; th e t a 2 := 0 .0 ; theta3 := 0.0; theta4 ;= 0.0; for i := 1 to size do begin thetal := thetal + table[i,i]; theta 2 := t h eta 2 + frowtot[i] * coltot[i]); theta3 := theta3 + ftable[i,i] * (rowtot[i] + coltot[i])); (* new version of theta4 *) for j := 1 to size do theta4 := t h e t a 4 + ( t a b l e [ i , j ] * s g r (rowtot[jl + c o l t o t [i ] ) ); end; thetal := thetal / n; t he t a 2 := th e t a 2 / sgr(n); theta3 := theta3 / sar(n); theta4 := theta4 / (sqr(n) * n); writeln; w r i t e l n ('Theta one = 1 ,t h e t a l :6 :5); writeln; wri teln('Theta two = 1 ,t he ta 2:6: 5); writeln; w r i t e l n ('Theta three = ',t h e t a 3 ;6 :5); writeln; writ eln f'Thet a four = ',theta4: 6 :5); (******* calculate khat and sigmahat *******) khat := (thetal - th e t a 2 ) / (1 - th e t a 2 ); 241 writeln; writeln; w r i t e l n ( 1K hat = *,khat:9:8); tem pi := temp2 := / ( sqr(ltemp3 := ( t h e t a l * (1 - t h e t a l ) ) / s a r ( 1 - t h e t a 2 ) ; 2 * ( ( 1 - t h e t a l ) * ( ( 2 * t h e t a l * t h e t a 2 ) - t h e t a 3 )) t h e t a 2 ) * ( l - t h e t a 2 ) ); ( sq r(1 - t h e t a l )* (th eta 4 -4 * sq r(th eta 2 ))) / s q r ( s q r (l-th eta 2 ) ); sigmahat := (tempi + temp2 + temp3) / n; writeln; writelni w r i t e l n ( 'Sigma hat = 1 /s i g m a h a t :9: 8 ): writeln; writeln; if (printan = 'y') or (printan = 'Y') then begin writeln(printer); w r i t e l n (p r i n t e r K hat = ,#khat:9:8); w r i t e l n (p r i n t e r ) ; w r i t e l n ( p r i n t e r S i g m a hat = ',s i g m a h a t : 9 :8 ); end; writeln('Normal program t e r m i n a t i o n ')j end. BIBLIOGRAPHY BIBLIOGRAPHY Aronoff/ S. 1981. Classification Accuracy - A Review. Technical Papers of the American Society of Photogrammet ry Fall Technical Meeting/ pp. 125-136/ American Society of P h o t o g r a m m e t r y / Falls Church/ Virginia. Benson/ A.S./ and S.D. DeGloria. 1985. Interpretation of Landsat-4 Thematic Mapper and Multispectral Scanner Data For Forest Surveys. Photogrammetric Engineering and Remote Sensing/ Vol. 51/ No. 9./ pp. 1281-1289. Bishop/ Y . / S. Fienberg/ and P. Holland. 1975. Discrete Multivariate Analysis - Theory and Practice. MIT Press/ Cambridge/ Massachusetts/ 575 p. Borden/ F . Y . / B.F. Merembeck/ D.N. Thompson/ B.J. Turner/ and D.L. Williams. 1974. Classification and Mapping of Coal Refuse/ Vegetative Cover Types/ and Forest Types By Digital Processing ERTS-1 Data. Proceedings Ninth International Symposium on Remote Sensing of Environment/ Ann Arbor/ Michigan/ pp. 133-152. Bryant/ E. 1983. Investigation of Forestry Resources and Other Remote Sensing Data. Final Technical Report NCC 5-22/ Dartmouth College/ Hanover/ New Hampshire. Bryant/ E . / A.G. Dodge/ Jr./ and S.D. Warren. I960. Landsat For Practical Forest Type Mapping: A Test Case. Photogrammetric Engineering and Remote Sensing/ Vol. 46/ No. 12/ pp. 1575-1584. Burgan/ R.E. and M.B. Shasby. 1984. Mapping Borad-Area Fire Potential From Digital Fuel/ Terrain/ and Weather Data. Journal of Forestry/ Vol. 82/ No. 4, pD. 228231. Chrisman/ N.R. 1982. Beyond Accuracy Assessment: Cor rec ­ tion of M i s c l a s s i f i c a t i o n . Proceedings of Fifth International Symposium on Computer-Assisted Cartography/ pp. 123-132. American Society of P h o t o g r a m m e t r y / Falls Church/ Virginia. Cohen/ J. 1960. A Coefficient of Agreement for Nominal Scales. Educational and Psychological Measurement/ Vol. 20/ No. 1/ pp. 37-46. Cohen/ J. 1968. Weighted Kappa: Nominal Scale Agreement with Provision For Scaled Disagreement or Partial Credit. Psychological Bulletin/ Vol. 70/ No. 4/ pp. 213-220. 242 243 Congalton, R.G. 1981. The Use of Discrete Multivariate Analysis Techniques for the Assessment of Landsat Classification Accuracy. Master thesis. Virginia Polytechnic Institute and State University, Blacksburg, Virginia. lllp. Congalton, R.G., J.T. Heinen, and R.G. Oderwald. 1983. Update and Review of Accuracy Assessment Techniques for Remotely Sensed Data. Remote Sensing Research Report 83-1, Virginia Polytechnic Institute and State University, Blacksburg, Virginia. Congalton, R . G . , and R.A. Mead. 1981. A Quantitative Method to Test for Similarity Between Photo Interpreters. Technical Papers of the American Society of Phot ogr ammetry 47th Annual Meeting, pp. 263-266, American Society of P h o t o g r a m m e t r y , Falls Church, Virginia. Congalton, R . G . , and R.A. Mead. 1983. A Quantitative Method to Test for Co nsi stency and Correctness in Photointerpretation. Photogrammetrie Engineering and Remote Sensing, Vol. 49, No. 1, pp. 69-74. Congalton, R . G . , R.A. Mead, R.G. Oderwald, and Heinen, 1981. Analysis of Forest Classification Accuracy. Remote Sensing Research Report 81-1, Virginia Polytechnic Institute and State University, Blacksburg, Virginia. Congalton, R.G., R.G. Oderwald, and R.A. Mead. 1982. Accuracy of Remotely Sensed Data: Sampling and Analysis Procedures. Remote Sensing Research Report 82-1, Virginia Polytechnic Institute and State University, Blacksburg, Virginia. Congalton, R.G. R.G. Oderwald, and R.A. Mead. 1983. Assessing Landsat Classification Accuracy Using Discrete Multivariate Analysis Statistical Techniques. Photogrammetric Engineering and Remote Sensing, Vol. 49, No. 12, pp. 1671-1678. Draeger, W.C. and D.M. Carneggie. 1974. Test Procedures for Remote-Sensing Data. Photogrammetrie Engineering Vol. 40, No. 2, pp. 175-181. Everitt, B.S. 1968. Moments of the Statistics Kappa and Weighted Kappa. The British Journal of Mathematical and Statistical Psychology, Vol. 21, Part 1, pp. 97103. 244 P i t z p a t r i c k - L i n s , K. 1981. Comparison of Sampling Procedures and Data Analysis for a Land-Use and LandCover Map. Photogra mmetric Engineering and Remote Sensing Vol. 47/ No. 3, pp. 343-351. Fleiss/ J.L. 1981. Statistical Methods for Rates and Proportions. John Wile y and Sons/ New York. 321 p. Fleiss/ J.L./ J. Cohen/ and B.S. Everitt, 1969. Large Sample Standard Errors of Kappa and Weighted Kappa. Psychological Bulletin, Vol. 72, No. 5, pp. 323-327. Fleming, M.D. and R.M. Hoffer. 1979. Machine Processing of Landsat MSS Data and DMA Topographic Data for Forest Cover Type Mapping. 1979 Machine Processing of Remotely Sensed Data Symposium, pp. 377-390, Purdue University, West Lafayette, Indiana. Fox, L. Ill, and K.E. Mayer. 1979. Using Guided Clustering Techniques for Mapping Forest Land Cover in Northern California. 1979 Machine Processing of Remotely Sensed Data Symposium, pp. 364-367, Purdue University, West Lafayette, Indiana. Fox, L. Ill, and K.E. Mayer. 1981. Using Ecological Zones to Increase the Detail of Landsat Classifications. Technical Papers of the American Society of Photogrammetry 1981 Fall Technical Meeting, pp. 113-124, Falls Church, Virginia. Fox, L. Ill, K.E. Mayer, and A.R. Forbes. 1983. Cla ssification of Forest Resources with Landsat Data. Journal of Forestry, Vol. 81, No. 5, pp. 283-287. Franklin, K.L., W.D. Hudson, and C.W. Ramm. 1983. Landsat Imagery for Identifying Coniferous Forest Types in Michigan. Research Report 448, Michigan State Unive rsi ty Agriculture Experiment Station, East Lansing, Michigan. Hahn, J.T. 1982. Timber Resource of Michigan's Southern Lower Peninsula, 1980. U.S. Forest Service, North Central Forest Experiment Station, Resource Bulletin NC- 6 6 . Hay, A.M. 1979. Sampling Designs to Test Land-Use Map Accuracy. Photogrammetric Engineering and Remote Sensing Vol. 45, No. 4, pp. 529-533. 245 Heller, R.C., Technical Coordinator. 1975. Evaluation of ERTS-1 Data For Forest and Rangeland Surveys. U.S. Forest Service, Pacific Southwest Forest and Range Experiment Station, Research Paper PSW-112. Hixson, M.M. 1981. Techniques for Evaluation of Aera Estimates. 1981 Machine Processing of Remotely Sensed Data Symposium, pp. 84-90, Purdue University, West Lafayette, Indiana. Hoffer, R.M., S.C. Noyer, and R.P. Mrocznski. 1978. A Comparison of Landsat and Forest Survey Estimates Forest Cover. Pro ceedings of the American Society of Photogram met ry Fall Technical Meeting, Albuquerque, New Mexico, pp. 221-231. Hord, R.M. and W. Brooner. 1976. Land-Use Map Accura cy Criteria. Photogra mmet ric Engineering and Remote Sensing Vol. 42, No. 5, pp. 671-677. Hudson, W.D. 1981. Density Slicing of Landsat Satellite Images for Forest Acreage Determination. Center for Remote Sensing, Michigan State University, East Lansing, Michigan. Hudson, W . D . , and K. Kittleson. 1978. Identification of Wood Energy Resources in Central Michigan. Remote Sensing Project, Michigan State University, East Lansing, Michigan. Jakes, P.J. 1982. Timber Resource of Michigan's Northern Lower Peninsula, 1980. U.S. Forest Service, North Central Forest Experiment Station, Resource Bulletin N C -6 2. Johnson, G.R., E.W. Barthmaier, T.W.D. Gregg, and R.E. Aulds. 1979. Forest Stand Classification in Western Washington Using Landsat Computer-Based Resource Data. Proceedings Thirteenth International Symposium on Remote Sensing of Environment, Ann Arbor, Michigan, pp. 1681-1696. Kalensky, Z. 1974. ERTS Thematic Map From Multidate Digital Images. Proceedings: Symposium on Remote Sensing and Photo Interpretation, International Society for Photogrammetry, Commission VII, Banff, Alberta, Canada. pp. 767-785. 246 Kalensky/ Z.D., W.C. Moore/ G.A. Campbell; D.A. Wilson; and A.J. Scott. 1979. Forest Statistics by ARIES Classification of Landsat Multispectral Images in Northern Canada. Proceedings Thirtheenth International Symposium on Remote Sensing of Environment; Ann Arbor; Michigan; pp. 789-811. Kalensky; Z.D.; W.C. Moore; G.A. Campbell; D.A. Wilson; and A.J. Scott. 1981. Summary Forest Resource Data from Landsat Images. Canadian Forestry Service; Petawawa National Forestry Institute; Chalk River; Ontario; Canada; Information Report PI-X-5. Kalensky; Z. and L.R. Scherk. 1975. Accuracy of Forest Mapping From Landsat Computer Compatible Tapes. Proceedings Tenth International Symposium on Remote Sensing of Environment; Ann Arbor; Michigan pp. 11591167. Kan; E.P. 1976. A New Computer Approach to Ma p Mixed Forest Features and Postprocess Multispectral Data. Proceedings of the American Society of Photogram metry 1976 Fall Convention; pp. 386-401; Falls Church; Virginia. Karteris; M .A . f W.P. Enslin; and J.Thiede. 1981. Aera Estimation of Forestlands in Southwestern Michigan From Landsat Imagery. Second Eastern Regional Remote Sensing Applications Conference; pp. 147-155; NASA Conference Publication 2198. Khorram; S. and E.F. Katibah. 1981. Use of Landsat Multspectral Scanner Data In Vegetation Mapping of a Forested Area. Technical Papers; American Society of Pho tog rammet ry 47th Annual Meeting; pp. 383-392; Falls Church; Virginia. Kirk; R.E. 1968. Experimental Design: Procedures for the Behavioral Sciences. Wadsworth Publishing Company; Belmont; California; 577 p. Krumpe; P.F.; J.D. Nichols; and D.T. Lauer. 1973. ERTS-1 Analysis of Wildland Resources Using Manual and Automatic Techniques. Proceedings Symposium Management and Utilization of Remote Sensing Data; American Society of Photogrammetry; pp. 50-66. Lillesand; T.M.; and R.W. Kiefer. 1979. Remote Sensing and Image Interpretation. John Wiley and Sons; New York; 612 p. 247 Mayer, K.E. and L. Fox XXI. 1981. Identification of Conifer Species Gr oupings from Landsat Digital Classifications. Photogrammetric Engineering and Remote Sensing, Vol. 48, No. 11, pp. 1607-1614. Mayer, K.E., L. Fox III, and J.L. Webster. 1979. Forest Condition Mapping of the Hoopa Valley Indian Reservation Using Landsat Data. Proceedings of Remote Sensing for Natural Resources, pp. 217-242, University of Idaho, Moscow, Idaho. Mazade, A.V. 1981. Ten-Ecosystem Study, Final Report. U.S. Forest Service, Nationwide Forestry Applications Program, Houston, Texas. LEMSCO-13491. Mead, R. and M. Meyer. 1977a. Landsat Digital Data Application to Forest Vegetation and Land-Use Classification in Minnesota. Remote Sensing Laboratory, University of Minnesota, St. Paul, Minnesota, IAFHE RSL Research Report 77-6. Mead, R . , and M. Meyer. 1977b. Landsat Digital Data Application to Forest Vegetation and Land Use Classification in Minnesota. 1977 Machine Processing of Remotely Sensed Data Symposium, pp. 270-279. Purdue University, West Lafayette, Indiana. Mead, R.A. and J. Szajgin. 1981. Landsat Classificatin Accuracy Assessment Procedures: An Account of a National Working Conference. 1981 Machine Processing of Remotely Sensed Data Symposium, pp. 202-204, Purdue University, West Lafayette, Indiana. Mead, R.A. and J. Szajgin. 1982. Landsat Classification Accuracy Assessment Procedures. Photogrammetric Engineering and Remote Sensing Vol. 48, No. 1, pp. 139-141. Michigan Inventory Advisory Committee, 1982. The Michigan Resource Inventory Act, Act 204. Land Resource Programs Division, Michigan Department of Natural Resources. Moik, J.G. 1980. Digital Processing of Remotely Sensed Images. National Aeronautics and Space Administration, Washington, D.C., NASA SP-431. Morrison, D.F. 1976. Multivariate Statistical Methods. McGraw-Hill Book Company, New York, 415 p. 248 Morrissey, L.A. and V.G. Ambrosia. 1982. Forestry Timber Typing Final Report, Tanana Demonstration Project. National Aeronautics and Space Administration, NASA Contractor Report 166391. Myers, W.L., G.R. Safir, A.L. A n d e r s o n , D.L. Mokma, E.P. Whiteside, H.A. Winters, R. Rieck, W.A. Mailia, J.E. Sarno, T.W. Wagner, J.T. Lewis, and J.E. Erickson. 1974. Use of ERTS Data for a Multi ­ di sci plinary Analysis of Michigan Resources Final Report. MSU Agriclutural Experiment Station Journal No. 7091, Michigan State University, East Lansing, Michigan. Nelson, R.F. and R.M. Hoffer. 1980. Procedure 1 and Forestland Classification Using Landsat Data. 1980. Machine Processing of Remotely Sensed Data Symposium, pp. 319-324, Purdue University, West Lafayette, Indiana. Newland, W . , D. Peterson, and S. Norman. 1980. Bulk Processing Techniques for Very Large Aeras: Landsat Classification of California. 1980 Machine Processing of Remotely Sensed Data Symposium pp. 306-317, Purdue University, West Lafayette, Indiana. NOAA. 1979. Climatological Data, Michigan. Vol. 94, No. 2. National Oceanic and Atmospheric Administr a­ tion, Environmental Data and Information Service, National Climatic Center, Asheville, North Carolina. Raile, G.K. and W.B. Smith. 1983. Michigan Forest Statistics, 1980. U.S. Forest Service, North Central Forest Experiment Station, Resource Bulletin NC-67. Rao, C.R. 1965. Linear Statistical Inference and Its Applications. John Wiley and Sons, New York, 522 p. Richards, J.A., D.A. Landgrebe, and P.H. Swain. 1982. Means for Utilizing Ancillary Information In Multispectral Classification. Remote Sensing of Environment, Vol. 12, pp. 463-477. A Roller N.E.G., and L. Visser. 1980. Accuracy of Landsat Forest Cover Type Mapping in the Lake States Region of the U.S. Proceedings Fourteenth International "Symposium on Remote Sensing of Environment, pp. 15111520, San Jose, Costa Rica. 249 Rose, G.A. 1978. A Comprehensive Inventory System for Forest Resource Management. Integrated Inventories of Renewable Natural Resources: Proceedings of the Workshop. U.S. Forest Service, Rocky Mountain Forest and Range Experiment Station, General Technical Report R M - 5 5 , pp. 463-468. Rosenfield, G.H., K. F i t z p a t r i c k - L i n s , and H.S. Ling. 1984. Sampling for Thematic Map Accuracy Testing. Photogrammetric Engineering and Remote Sensing Vol. No. 1, pp. 131-137. 48, Rosenfield, G.H., M.L. Melley. 1980. Applications of Statistics to Thematic Mapping. Photogrammetric Engineering and Remote Sensing Vol. 46, No. 10, pp. 1287-1294. Salazar, L.A. 1982. Remote Sensing Techniques Aid In Preattack Planning for Fire Management. U.S. Forest Service, Pacific Southwest Forest and Range Experiment Station, Research Paper PSW-162. S c h o w e n g e r d t , R.A. 1983. Techniques for Image Processing and Classification in Remote Sensing. Academic Press, New York, 249 p. Searle, S.R. 1966. Matrix Algebra For the Biological Sciences (Including Applications in Statistics). John Wiley and Sons, New York. 296 p. Shasby, M.B., R.E. B u r g a n , and G.R. Johnson. 1981. Broad Area Fuels and Topogr aphy Mapping Using Digital Landsat and Terrain Data. 1981 Machine Processing of Remotely Sensed Data Symposium, pp. 529-537, Purdue University, West Lafayette, Indiana. Slater, P.N. 1979. A Re-examination of the Landsat MSS. Photogrammetric Engineering and Remote Sensing. Vol. 45, No. 11, pp. 1479-1485. Smith, W.B. 1982. Timber Resource of Michigan's Eastern Upper Peninsula, 1980. U.S. Forest Service, North Central Forest Experiment Station, Resource Bulletin NC-64. Snedecor, G.W., and W.G. Cochran. 1967. Statistical Methods. The Iowa State University Press, Ames, Sokal, R.R., and F.J. Rohlf. and Co., San Francisco. 1969. Biometry. 776 p. W.H. Iowa. Freeman 250 Spencer/ J.S. Jr. 1982. Timber Resource of Michigan's Western Upper Peninsula/ 1980. U.S. Forest Service/ North Central Forest Experiment Station/ Resource Bulletin NC-60. Spencer/ J.S. Jr. 1983. Michigan's Fourth Forest Inventory: Area. U.S. Forest Service/ North Central Forest Experiment Station/ Resource Bulletin N C - 6 8 . Strahler/ A.H./ J.E. Estes/ P.F. Maynard/ F.C. Hertz/ and D.A. Stow. 1980. Incorporating Collateral Data In Landsat Classification and Modeling Procedures. Proceedings Fourteenth International Symposium on Remote Sensing of Environment/ pp. 10091026/ San Jose/ Costa Rica. Strahler/ A.H./ T.L. Logan and N.A. Bryant/ 1978. Improving Forest Cover Classification Acc uracy From Landsat By Incorporating Topographic Information. Proceedings Twelfth International Symposium on Remote Sensing of Environment/ pp. 927-942/ Ann A r b o r f Michigan. Strahler/ A.H./ T.L. Logan/ and C.E. Woodcock. 1979. Forest Classification and Inventory System Using Landsat/ Digital Terrain/ and Ground Sample Data. Proceedings Thirteenth International Symposium on Remote Sensing of Environment/ pp. 1541-1557/ Ann Arbor/ Michigan. Strommen/ N.D. 1968. Michigan Snowfall Statistics: First 1-/ 3-/ 6 -/ 12-inch Depths. Michigan Weather Service/ Michigan Department of Agriculture/ L a n s i n g f Michigan. Strommen/ N.D. 1974. Michigan Snow Depths. Michigan Weather Service/ Michigan Department of Agriculture/ Lansing/ Michigan. Swain/ P . H . / and S.M. Davis (eds.). 1978. Remote Sensing: The Quantitative Approach. McGraw-Hill/ New York/ 396p. Taranik/ J.V. 1978. Cha racteristics of the Landsat Multispectral Data System. U.S. Geological Survey Open-File Report 78-187/ Sioux Falls/ South Dakota. Turk/ G. 1979. GT Index: A Measure of the Success of Prediction. Remote Sensing of Environment Vol. 8 f pp. 65-75. U.S. Forest Service. 1983. Automated Forest Classification and Inventory in the Eldorado National Forest. USDA/ Forest Service/ Region 5 Contract No. 53-9158-6504. 251 U.S. U.S. Geological Survey. 1977. EPOS Digital Image Enhancement System (EDIES) "Fact Sheet." EROS Center/ Sioux Falls/ South Dakota. Data Geological Survey. 1979. Landsat Data Users Handbook. Revised Edition. U.S. Geological Survey/ EROS Data Center/ Sioux Falls/ South Dakota. Van G e n d e r e n f J.L./ B.F. Lock/ and P.A. Vass. 1977. Testing the Acc uracy of Remote Sensing Land Use Maps. Proceedings of the Eleventh International Symposium on Remote Sensing of Environment/ pp. 615-623/ Environ­ mental Research Institute of Michigan/ Ann Arbor/ Michigan. Walsh/ S.J. 1980. Landsat Data. pp. 11-26. Coniferous Tree Species Mapping Using Remote Sensing of Environment/ Vol. 9/ Werth/ L.F. 1981. An Evaluation of ISOCLS and CLASSY Clustering Algorithms for Forest Classification in Northern Idaho. 1981 Machine Processing of Remotely Sensed Data Symposium/ pp. 11-17. Purdue University/ West Lafayette/ Indiana. Williams/ D.L. 1976. A Can opy-Related Stratification of a Southern Pine Forest Using Landsat Digital Data. Pr oceedings of the American Society of Photogram met ry Fall Convention/ pp. 231-239/ American Society of Photogrammetry/ Falls Church/ Virginia. Woodcock/ C.E./ A.H. Strahler/ and T.L. Logan. 1980. Stratification of Forest Vegetation for Timber Inventory Using Landsat and Collateral Data. Proceedings Fourteenth International Symposium on Remote Sensing of Environment/ pp. 1769-1787/ San Jose/ Costa Rica.