INFORMATION TO USERS This was produced from a copy of a document sent to us for microfilming. While die most advanced technological means to photograph and reproduce this document have been used, the quality is heavily dependent upon the quality of the material submitted. The following explanation of techniques is provided to help you understand markings or notations which may appear on this reproduction. 1.The sign or “target” for pages apparently lacking from the document photographed is “Missing Page(s)” . If it was possible to obtain the missing page(s) or section, they are spliced into the Him along with adjacent pages. This may have necessitated cutting through an image and duplicating adjacent pages to assure you of complete continuity. 2. When an image on the film is obliterated with a round black mark it is an indication that the film inspector noticed either blurred copy because of movement during exposure, or duplicate copy. Unless we meant to delete copyrighted materials that should not have been filmed, you will find a good image of the page in the adjacent frame. 3. When a map, drawing or chart, etc., is part of the material being photo­ graphed the photographer has followed a definite method in “sectioning” the material. It is customary to begin filming at the upper left hand comer of a large sheet and to continue from left to right in equal sections with small overlaps. If necessary, sectioning is continued again-beginning below the first row and continuing on until complete. 4. For any illustrations that cannot be reproduced satisfactorily by xerography, photographic prints can be purchased at additional cost and tipped into your xerographic copy. Requests can be made to our Dissertations Customer Services Department. 5. Some pages in any document may have indistinct print. In all cases we have filmed the best available copy. University' Microfilms International 3 00 N. ZEEB ROAD, ANN ARBOR. Ml 48 1 0 6 18 BED FORD ROW, LONDON WC1R 4E J, ENGLAND 8112099 K a r t e r is , M ic h a el A po st o l o s AN EVALUATION OF SATELLITE DATA FOR ESTIMATING THE AREA OF SMALL FORESTLANDS IN THE SOUTHERN LOWER PENINSULA OF MICHIGAN Ph.D. Michigan State University University Microfilms International 300 N. Zeeb Road, Ana Arbor, MI 48106 1980 PLEASE NOTE: In a ll cases t h i s material has been film ed 1n the b est p o ssib le way from the a v a ila b le copy. Problems encountered with th is document have been id e n tifie d here with a check mark . 1. Glossy photographs 2. Colored I llu s tr a tio n s 3. Photographs with dark background 4. Illu str a tio n s are poor co p y _______ 5. ° r in t shows through as there 1s te x t on bothsid es o f page _ _ _ _ _ 6 . I n d is tin c t, 1/ ✓ broken or small p rin t onseveral pages iS 7. T ightly bound copy w ith p rin t lo s t in spine ________ 8. Computer printout pages with In d istin c t p rin t _ _ _ 9. Page(s) — _ _ _ _ _ lacking when m aterial receiv ed , and not a v a ila b le from school or author 10. Page(s) _________ seem to be m issing in numbering only as te x t follow s 11. Poor carbon copy _ _ _ _ _ _ 12. Not orig in a l copy, several pageswith blurred 13. Appendix pages are poor copy 14. Original copy with lig h t t y p e _________ 15. Curling and wrinkled p a g e s ________ 16. Other type _____ ______ ________________________________________ _ _ University M icrcxilms International 3 0 0 N. Z EE B RD.. ANN ARBOR. Ml 4 8 1 0 6 ( 3131 761 4 7 0 0 AN EVALUATION OF SATELLITE DATA FOR ESTIMATING THE AREA OF SMALL FORESTLANDS IN THE SOUTHERN LOWER PENINSULA OF MICHIGAN By Michael Apostolos Karteris A DISSERTATION Submitted to Michigan State University in partial fulfillment of the requirements for the degree of DOCTOR OF PHILOSOPHY Department of Forestry 1980 ABSTRACT AN EVALUATION OF SATELLITE DATA FOR ESTIMATING THE AREA OF SMALL FORESTLANDS IN THE SOUTHERN LOWER PENINSULA OF MICHIGAN By Michael Apostolos Karteris This study evaluated the use and potential of Landsat images for mapping and estimating acreage of small scattered forestlands in Barry County, Michigan. Four Landsat images, two from late winter and two from early fall, were tested. Three of these images, the winter black-and-white band 5, the winter color and the fall color, were employed without modification. The fourth image was a diazo color composite of the fall scene produced by manipulating a black-andwhite transparency of each Landsat band to enhance the appearance of the forest resources. Forestlands as small as 2.5 acres were mapped from each Landsat data source. The maps for each image were then compared with a detailed forest type map that had been pre­ viously prepared by another investigator using color infrared photography and extensive ground truth information. Compari­ son of the Landsat maps with this source detected mapping errors, which were categorized as those of commission and omission, and then further classified into boundary or iden­ tification errors. The commission errors were usually higher than those of omission, and most of the error cases were boundary errors of less than five acres. The most frequently misclassified areas were agriculture lands, treed-bogs, brushlands, and lowland and mixed hardwood stands, while stocking level affected interpretation more than stand size. The overall level of the interpretation performance was ex­ pressed three ways, through the estimation of classification, interpretation and mapping accuracies. These accuracies ranged between 74 and 98 percent. The overall recommendation of the study is that, con­ sidering errors, accuracy, and cost, winter color imagery is the best Landsat alternative for mapping small forest tracts. However, since the availability of cloud-free winter images of the study area is significantly lower than images for other seasons, a diazo enhanced image of a fall scene is recommended as the next best alternative. To My Parents# Apostolos and Mersina? My Wife, Kaity and My Son, Apostolos ACKNOWLEDGMENTS It would have been impossible to conduct a study of this nature without the assistance, cooperation and under­ standing support of many individuals. The encouragement and invaluable guidance of the Doctoral Committee remains the salient contributory factor toward the development of the project. The writer would like to express his appreciation to Dr. Carl Ramm, Chairman of the committee; Mr. William Enslin, thesis director; Dr. Melvin Koelling, Dr. Victor Rudolph and Dr. Charles Olson, members of the committee, for exposing him to the benefits of their educational experience and scientific exactness. The writer is especially grateful to Dr. Carl Ramm who was an invaluable source of advice, assistance and en­ couragement. He was instrumental in guiding the writer's efforts and provided instructive criticism when it was needed. Mr. William Enslin, Manager of the Center for Remote Sensing, went far beyond the call of duty as a member of the guidance committee in giving assistance. - The numerous sessions spent with him were particularly helpful in for­ mulating the theoretical and practical framework of this iii study. His guidance, encouragement, insightful suggestions and advice, patience, compassion and friendship have set an example which the writer will continually seek to emulate personally and professionally. Mr. Enslin created a favor­ able environment for research and helped make the writer's stay at the Center so rewarding. The writer deeply acknowledges the sincere contribu­ tion and cooperation of other members of the Center for Remote Sensing for their constructive counseling. Especially he would like to thank Mr. David Lusch for his willingness to lend his technical knowledge, experience and service to the research; Mr. Richard Hill-Rowley for his constructive comments and help throughout this study and Miss Robin Landfear for her assistance in doing most of the cumbersome calculations and drawing all the graphs in the study. Particular gratitude is owed to Miss Elizabeth Bartels, secretary of the Center for Remote Sensing, for her incredible patience in typing the early drafts and for her excellent work in typing the final draft of the thesis. The writer is most grateful to Dr. George Bouyoucos, Emeritus Professor of Michigan State University who kindly granted him a fellowship which made it possible to attend Michigan State University. For financial assistance throughout this study the writer wishes to express appreciation t;o the Michigan Depart­ ment of Natural Resources (Forest Management Division) and the National Aeronautics and Space Administration. The study could not have been completed without the absolute understanding, patience and sacrifice of my wife, Kaity, and my son, Apostolos, who provided much of the in­ centive and strong and continuous encouragement needed in order to make this work possible. Without the kindly and untiring help of the above men­ tioned, the realization of this manuscript would be very difficult. v TABLE OF CONTENTS Page LIST OF T A B L E S ..................................... ix LIST OF F I G U R E S ................................... xi Chapter I. INTRODUCTION A. B. C. ............................ 1 General . . . . : ................... Landsat System........................ Objectives ....................... 1 3 5 II. LITERATURE REVIEW ....................... 8 III. THE STUDY A R E A ......................... 20 A. B. 20 IV. V. G e n e r a l .............................. Major Forest Cover Types in Barry C o u n t y ........................... 23 DATA A C Q U I S I T I O N ....................... 28 A. B. 28 35 Landsat D a t a ....................... Reference D a t a ..................... STUDY MATERIALS ,DESIGN AND DATA COLLEC­ TION 37 A. 37 Study M a t e r i a l s ..................... 1. 2. Standard Images: winter blackand-white; winter color; fall c o l o r ....................... Diazo ColorComposite ........... vi 37 41 Chapter Page B. Study Design and Data Collection 1. 2. 3. VI. Selection of the Training Set . . Interpretation Procedures . . . . Verification and Error Types . . RESULTS AND D I S C U S S I O N ................... A. Interpretation Errors B. 61 Commission Errors .............. Omission E r r o r s ................. Commission and Omission Errors-Combined Effects ............. 63 70 Accuracy Analysis .................... 75 Classification Agreement . . . . Interpretation Agreement . . . . Mapping Agreement.. .............. Size of Interpretation Errors . . . . 1. 2. D. 46 48 49 61 1. 2. 3. C. 46 .............. 1. 2. 3. 73 76 78 83 84 Commission Errors .............. Omission E r r o r s ................. 86 93 Land Cover/Use Categorization of the Interpretation Errors ........... 98 1. 2. 3. Commission Errors.. .............. 99 Omission E r r o r s .................... 104 Effects of Silvicultural Condi­ tion on Omission Errors . . . 108 E. Additional Results 1. 2. 3. 4. VII. . . .................. 110 Threshold S i z e ....................110 C o s t s .............................. Ill Forest Tracts Mapped Only From Landsat Images.. .............. 118 Availability of Landsat Images and Cloud Cover Restrictions . 119 SUMMARY AND C O N C L U S I O N S ................... 123 B I B L I O G R A P H Y ........................................ 128 Page APPENDICES Appendix A. LANDSAT PARAMETERS AND CHARACTERISTICS . . 139 B. ADDITIONAL STUDIES OP LANDSAT USES IN F O R E S T R Y ................................ 144 C. DIAZO PROCESSING AND METHODOLOGY ......... 146 D. INTERPRETATION EQUIPMENT ................. 154 E. TABULATION OF THE R E S U L T S ................. 156 viii LIST OF TABLES Table 1. 2. Page Characteristics of the selected Landsat s c e n e s ....................................... Forest vegetation and land cover/use classi­ fication system developed for qualitative evaluation of the interpretation errors . 35 53 3. Forest stand size and stocking level cate­ gories and brush (upland, lowland) ground cover c l a s s i f i c a t i o n s ................ 54 4. An example of a commission error data sheet for color winter scene of Johnstown Town­ ship ....................................... 58 An example of an omission error data sheet for winter color scene of Johnstown Town­ ship ....................................... 59 Errors of interpretation performance by type of Landsat image for Barry County, Michi­ gan ....................................... 64 Classification, interpretation and mapping agreement by type of Landsat image. . . . 77 5. 6. 7. 8. Time spent on visual interpretation operations by Landsat i m a g e ........................ 114 9. Forestlands mapped from Landsat images but not included in the reference d a t a .............. 118 10. Available Landsat images of Barry County, Michigan (path 23 and row 30) by percent cloud cover and s e a s o n ...................... 121 E— 1. Interpretation errors of Barry County by town-. ship for black-and-white image of the winter scene (February 26, 1 9 7 9 ) ............... 156 V— xx Page Table E-2. Interpretation errors of Barry County by town­ ship for false color composite of the fall scene (September 12, 19 7 9 ) ............... 158 E-3. Interpretation errors of Barry County by town­ ship for false color composite of the winter scene (February 26, 1 9 7 9 ) ............... 160 E-4. Interpretation errors of Barry County by town­ ship for diazo color composite of the fall scene (September 12, 1 9 7 9 ) ............... 162 E-5. Total commission error by size class (in acres) and type of Landsat i m a g e ..................164 E-6. Boundary commission errors by size class (in acres) and type of Landsat i m a g e ........... 166 E-7. Identification commission error by size class (in acres) and type of Landsat image . . . 168 E-8. Total omission error by size class and type of Landsat image ............................ 170 E- 9. Boundary omission error by size class (in acres) and type of Landsat i m a g e ........... 172 E-10. Identification omission error by size class (in acres) and Landsat i m a g e ............... 174 E-ll. Commission errors by land cover/use categories and type of Landsat i m a g e ..................176 E-12. Omission errors by forest cover type and Land sat image ................................ 178 E-13. Stand size and stocking level codes ......... 180 E-14. Omission errors by forest type stocking level classes and types of Landsat image . . . . 181 Omission errors by forest type stand size classes and types of Landsat image . . . . 183 E-15. x LIST OF FIGURES Figure 1. 2. Page Barry County, Michigan, showing its sub­ division into 16 townships and location in the s t a t e ............................... Position of the study area within the Landsat scene of path 23 and row 30 of the World­ wide Reference System and location of the centers of the scenes required to cover the whole s t a t e ........................ 21 31 3. Barry County study area delineated on a false color composite scene imaged by Landsat-3, September 12, 1979 ................... 4. Landsat black-and-white band 5 subimage of Barry County, Michigan taken on February 26, 1979 ............................. 39 Landsat false-color composite subimage of Barry County, Michigan taken on February 26, 1979 ................................. 40 Landsat diazo false-color composite subimage of Barry County, Michigan taken on Sep­ tember 12, 1979 44 Forest interpretation map from the color fall image of Woodland Township, Barry County, showing commission and omission errors for different forest types ................... 56 Interpretation commission and omission errors expressed as percentages of the total forest acreage in the study area by type of Landsat i m a g e ......................... 65 Percent boundary and identification commis­ sion errors by type of Landsatimage . . . 69 Percent boundary and identification omission errors by type of Landsat i m a g e ......... 72 5. 6. 7. 8. 9. 10. xi 33 Figure 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. Page Variation of the estimated total forest area (in acres) by type of Landsat image ... Percent classification agreement by type of Landsat i m a g e .............................. 74 79 Percent interpretation agreement by type of Landsat i m a g e ......................... 82 Percent mapping agreement by type of Landsat i m a g e ...................................... 85 Distribution of total commission error by size class (acres) and type of Landsat i m a g e ...................................... 88 Distribution of boundary commission error by size class (acres) and type of Landsat i m a g e ................................. 90 Distribution of idnetification commission error by size class (acres) and type of Landsat i m a g e ...................................... 91 Distribution of total omission error by size class (acres) and type of Landsat i m a g e ...................................... 94 Distribution of boundary omission error by size class (acres) and type of Landsat i m a g e ...................................... 96 Distribution of identification omission error by size class (acres) and type of Landsat i m a g e ...................................... 9 21. Commission error acreage by land cover/use category or subcategory and type of Land­ sat i m a g e ...................................102 22. Acreage percent omission error (acres) by forest cover type category and Landsat i m a g e ....................................... 107 23. Smallest size forest tracts visually depicted on the various types of Landsat images . . C-l. Characteristic curves of a' yellow diazo film at various exposure values ................ xii 112 148 7 Page Figure C-2. Graphical presentation of a hypothetical rela tion between the density range of scene elements in a Landsat image and the correS' ponding stretched or compressed density range on a high contrast and low contrast diazo copy of the image, respectively . . xiii 150 CHAPTER I INTRODUCTION A. General Sound management of forest resources entails three phases of activities: 1) resources inventory, 2) the development of improved plans or programs and 3) the adaptation and applica­ tion of these programs. A complete and detailed selection of current data during the inventory process is the primary input and basic prere­ quisite for a justifiable forest management decision-making process. A well designed and intelligent inventory system provides valuable, extensive, quantified and qualified information on the forest to be managed. Information such as forest location, acreage, volume, type and stand condition, land productivity, areas suitable for new plantations, geological characteristics of the area, surface and subsur­ face water and ownership should be included in the inventory output which will be considered by the forest manager. On the other hand, forestlands (especially small private tracts) are continuously shifted to other uses, such as agriculture. Also, other more rapid changes in the forest environment are caused naturally by fires, insect and/or disease epidemics and by cultural and technological developments such as air pollution. To keep track of all these changes, and to update the information about the condition of forest resources, requires repeated inventories. All this information will be used by the forest manager to amend managerial programs, to improve management practices and policies and/or adjust the regional and local programs according to policy changes. A sound inventory system requires efficient and costeffective data collection techniques and methods. It is the forest manager's responsibility to investigate, adopt and utilize those which best fit the conditions of his forest area and to give the required information. Forest inventories using conventional techniques and methods are the best way of acquiring the desired information, but they are very expen­ sive and time-consuming procedures. Such considerations become important with budget constraints; however, data should be kept up-to-date by taking new inventories at short inter­ vals {perhaps five years) even though costs continue to increase. Remote sensing techniques — especially vertical aerial photography using various film types and scales — integral part of many forest activities. are an This is due to high image quality, the development of photointerpretation and photogrammetric techniques, the availability of qualitative and quantitative keys and tables and the forester's familiar­ ity and acceptance of the new element. Most importantly, aerial photography can be used as a tool for gathering information about forest resources, increasing the accuracy , * at a given cost and/or achieving the predetermined level of accuracy with less labor and time. However, aerial photography cannot completely replace field work in forest inventories. Only a combination of photo and ground information will give the desired results. Furthermore, the dynamic state of the forest environment and the extended acreage of the forest resources require a synoptic view on a repetitive basis at short intervals. With an airborne photographic system it is almost impossible to fulfill these requirements because of the high cost and effort involved. B. Landsat System A new dimension in remotely acquired information (syn­ optic view, repetitive coverage, multispectral imagery, digital data format) about the earth's resources and their changes over time was initiated with the Landsat (Land Satellite) program in July of 1972. At that time NASA (National Aeronautics and Space Administration) launched the first unmanned satellite, ERTS-1 (Earth Resources Technology Satellite-1) which was later named Landsat-1. Since then, two more satellites identical to Landsat-1 have been launched. They are the only ones operational to date. All these satellites have been specifically designed on an experimental basis to collect information on the various features and conditions of the surface of the earth from an altitude of 570 miles (920 kilometers). The primary remote sensing system on board the satel­ lites, the Multispectral Scanning System (MSS), records the reflected radiation from an area of 79 x 79 meters of the earth's surface in our different spectral bands, two in the visible and two in the reflected infrared portion of the electromagnetic spectrum. To improve the quality of the received MSS data, several operations (e.g., radiometric and geometric corrections and edge and contrast enhancement techniques) are applied. The processed data are available 2 3 in a photo-like and digital format from the EROS Data 4 Center. Landsat technology appears to offer an alternative approach to solve some forest inventory problems. Synoptic view, repetitive coverage and multispectral imagery are desired characteristics of data employed for forestry pur­ poses. However, several questions have been raised concerning Electromagnetic spectrum is a series of electromagnetic radiations arranged according to wavelength or frequency. The spectrum extends from the shortest cosmic rays through the visible and infrared radiation to the radio energy. 2 Photo-like data are images produced from the digital satellite data by applying certain procedures at the EROS Data Center. They are available in print and transparency (positive-negative format). 3 Digital data are whole integers representing the brightness values of the picture elements. They range from 0 to 127 for bands 4, 5 and 6 and 0 to 63 for band 7 and they are recorded on computer compatible tapes (CCT's). 4 EROS Data Center, Sioux Falls, South Dakota 57198. Landsat by forest managers and scientists: 1} what informa­ tion does Landsat imagery contain for forestry; 2) how much of this information can be extracted; 3) what level of accuracy can be reached when extracting this information; and 4) what is the cost? Since the introduction of the Landsat program, forest scientists have been involved in various investigations try­ ing to answer ,these four basic questions. However, the problem is very complicated as many factors are involved (i.e., season of acquiring data, vegetation types, geomorphol­ ogy/ ground resolution of the system, radiometric resolution). Because of this, more research is needed in order to deter­ mine the potentials and drawbacks of the Landsat system. The Michigan Department of Natural Resources, Forest Manage­ ment Division, has posed the same general questions. The Department wants to know the efficiency of employing the new technology in various forest activities in the state, and the requirements in terms of personnel, experience and equipment to utilize Landsat data. This study was conducted to answer some of the questions and construct a comprehensive framework for using Landsat technology for forestry in Michigan. C. Objectives This study assessed the use of spaceborne remote sensing technology to detect, identify and delineate scattered forestlands. It examined various types of Landsat images to determine the extent to which they could provide the desired information to the forest manager. Of main interest in this study was the inventory of scattered small forest­ lands. The difference between this type of spatial distribu­ tion and an extensive contiguous forest of the same acreage is that, in the case of scattered small forestlands, the ratio of total length of tract boundaries to the actual forest acreage is very high. In terms of photointerpretation and mapping, this situation means that the interpreter has to detect, locate and trace many more boundaries which implies a higher probability of making locational errors. This, in turn, increases the probability of making classification errors negatively affecting the overall interpretation per­ formance and utility of the system. The specific objectives of the study were to: 1. Develop a methodology for manual interpretation of scattered small forestlands on Landsat images for two alternative seasons. 2. Assess the accuracy of the interpretation performance. 3. Analyze and evaluate the various interpretation errors qualitatively and quantitatively. 4. Evaluate the effect of the seasonal factor on the interpretation. 5. Evaluate Landsat based technology and procedures in terms of time, cost, personnel and equipment requirements for forest inventory purposes in Michigan. Based on the information obtained and conclusions drawn from this study, the Michigan Department of Natural Resources, Forest Management Division, will decide whether Landsat tech­ nology is a reliable and promising alternative to be inte­ grated into the periodic inventories of the state's forest resources'. CHAPTER II LITERATURE REVIEW Since the initiation of the Landsat program eight years ago (1972), professional foresters have expressed considerable interest in this new technology. Many investigations have been conducted to evaluate the capability and disadvantages of Landsat data. Special consideration has been given to spatial,6 spectral6 and radiometric7 resolutions and other characteristics in an attempt to solve certain forestry problems. Researchers have been trying to determine the efficiency with which MSS data provides necessary information to the forest manager. In an effort to accurately determine the degree of detail which can be extracted from Landsat data, 5 Spatial resolution is a measure of the smallest identi­ fiable feature. In Landsat it is called a "pixel" (for picture element) and it covers an area of 56 x 79 meters or 1.1 acres- T0.44 hectares). g Spectral resolution is a measure of the discreteness of the spectral band widths (spectral band width is an inter­ val in the electromagnetic spectrum which is defined by two wavelengths). Three of Landsat's sensors are sensitive over 0.2 pm wavelength range, whereas the fourth one is sensitive over 0.3 pm range. 7 Radiometric resolution is the sensitivity of the sensor to distinguish between gray levels. Three of Landsat's sensors can distinguish among 128 gray levels, whereas the fourth one can distinguish among 64 gray levels. 8 several individuals have conducted many studies covering a whole range of forest activities. In many instances emphasis has been placed on obtaining a quantitative expression (percentage) of estimation accuracy by comparing findings with available ground observations. These studies have entailed either manual interpretation of Landsat imagery, which is similar to conventional photointerpretation of aerial photography, or to more sophisticated computer-assisted analysis and classification of digital Landsat data. A few studies have looked at both. Studies were conducted to investigate, evaluate and demonstrate the feasibility, advantages and disadvantages of using Landsat technology in detecting, identifying, locating and mapping forest resources. Findings of the more significant findings are highlighted as follows. Tueller et al. (1973) reported that the minimum density for identifying the pinyon/juniper ecotone on Landsat-1, false color composite imagery, taken in July, was as low as 16.4 trees per acre provided the trees were close and con­ tinuous. Furthermore, areas of 55-60 acres with an acreage density of 30 trees or more per acre could be easily identi­ fied. In a computer-assisted analysis of Landsat-1 data taken in August to classify agricultural crops and forests in Michigan, Safir et al. (1973) found that forests could be classified correctly about 85 percent of the time. Lauer et al. (1973) noted that a skilled interpreter 10 could identify and distinguish forestland from non-forestland with more than 86 percent accuracy on Landsat-1 false color composite imagery. summer They also reported that this interpretation was done nearly 20 times faster in comparison to the use of black-and-white photography at a scale of 1:15,840. Erb (1973) reported that pine timber stands of ten acres and larger in size could be detected visually on black-andwhite and color composite images, provided that contrast with the surrounding area was good. Furthermore, the area measurement of large forest stands (on classification maps), using supervised computer-assisted classification, g differed by 11 percent from corresponding data selected from aircraft photography. Benson and Lauer (1973) conducted a series of quantita­ tive, manual interpretation tests on various image types (black-and-white, false color, enhanced color, etc.), taken in late August to determine which image provided the maximum amount of information for an optimum forest classification. The three-band (4, 5, 7) digitally enhanced images and twoband (5, 7) photographically made color composites gave the Q Computer-assisted classification is a computer-imple­ mented process of assigning individual pixels of a multispectral imagery to a class according to a pre-specified rule. If the classes have been defined based on inherent data characteristics, the classification is called unsupervised, whereas if the classes have been defined based on representa­ tive training areas of known characteristics, the classifica­ tion is called supervised. 11 best results. Classification with accuracies of 52 and 50 percent respectively were obtained. Hoffer et al. (1974) used computer-assisted techniques to identify and map forest cover types in mountainous areas of southwestern Colorado, from Landsat-1 digital data. An overall correct classification performance for coniferous and deciduous trees of about 94.3 percent was obtained. Results obtained from the use of only two channels and all four channels were almost the same. (5 and 7) Furthermore, acreage estimates of the various forest cover types with the actual values yielded a correlation coefficient of 0.982 at the 0.9 5 probability level. Heath (1974) investigated the applicability of various image and digital Landsat data to classify timber types on the Sam Houston National Forest in Texas. He reported that computer-assisted analysis gave better mapping results than manual interpretation 74 percent of the time. Lee (1974) tested the suitability of August and September Landsat imagery for monitoring forest management operations. Detailed mapping was difficult because of the relatively low resolution of the system. The minimum area delineated was ten acres, although this depended on the contrast with the surroundings. It was pointed out that Landsat data are best for broad forest resource inventory and extensive management activities. Jobin and Beaubien (1974) mapped broad vegetation cover types by a manual interpretation of Spring (April 2) and 12 Pall (October 4) Landsat imagery. They reported that satellite imagery is a promising tool, especially if photographic or computer enhancement techniques are employed. Eller and Ulliman (1974) concluded that maps of coniferhardwood delineations, based on manual interpretation of a band 5 winter scene (snow covered), were not accurate. Attempts to interpret forested areas from a combination of winter and summer scenes using an additive color viewer with color coded density level slice analysis of selected scenes did provide significant improvement. Aldred (1974) discussed how an experimental design should be planned to determine the reliability of Landsat imagery for interpretation and mapping of forest resources. Accuracy and cost should be the bases for evaluating various methods of extracting information from Landsat data. A preliminary trial of operator/band combinations indicated, that manual interpretations of bands 5 and 6, combined and photographically enhanced by an experienced interpreter, gave the best results (6.68 mean square error of the areal proportions between trial and standard). The interpretation of a standard false color composite print gave a mean square error of 7.28. Lee (1975a, 1975b) evaluated the reliability of manually identifying and measuring clear-cut areas on September 1972 and August 1973 imagery. He used a multidate, enhanced color print, a black-and-white band 5 of the September scene and a computer-compatible tape (CCT) of the August scene. The 13 interpretation of band 5 and the computer-assisted analysis were not successful because the forest boundaries were not always clear. On the other hand, interpretation and mapping using color prints was more successful. Kan and Dillman (1975) reported that temporal analysis, using early and late spring Landsat data, improved by up to 11 percent the classification accuracy of forests as opposed to single-season analysis. Furthermore, they found that two channel analyses were more effective, performing as well as the three or four channels for simple seasonal or temporal analysis. Kalensky and Scherk (197 5) examined the applicability of Landsat data in computer-compatible tape format for forest mapping. They found that for single-date imagery, the September 5 imagery yielded the best overall classification (81 percent) and mapping (72 percent) accuracies, whereas the March 1 imagery was least satisfactory. Classification based on channels 5 and 7 gave the same accuracies as when all four channels were used. Multidate image processing increased only marginally the classification and mapping accuracies; however, the pictorial outputs had a better dis­ play quality. Messmore et al. (1975) conducted a study to evaluate the forest mapping capability of summer (August 30) Landsat-1 MSS digital data. Using an unsupervised technique, followed . 14 by a supervised maximum likelihood classification g they found that training data were classified with an accuracy of 95 percent based entirely on the analysis of channels 6 and 7. Lee (1976) evaluated the feasibility of Landsat digital data (August 12, July 20) for forestland classification. The unsupervised classification was very poor; however, a supervised one was very successful. He noted that enlarge­ ment of each pixel four times gave more information but was probably not economical because of the computer time in­ volved. He recommended that operational information regard­ ing logged and burned-over areas can be updated using the Image 100 computer s y s t e m ^ or photographic enhancement techniques. He also pointed out that pixel by pixel classifi­ cation is suitable as the first stage in a multistage forest sampling design. Santos et al. (197 5) tried to verify the capability of the Landsat system as a means of monitoring deforestation in the Amazon region. Manual interpretation and computer-assisted g Supervised maximum likelihood classification is based on the performance of the maximum likelihood rule which quantitatively evaluates both the variance and correlation of the spectral response patterns of each preselected training set when it classifies an unknown pixel. Image 100 computer system is an interactive image analysis system designed for classifying multi-spectral scanner data. It was developed by the General Electric Company and consists of both hardware and proprietary software. Trade names, trademarks or commercial enterprises or pro­ ducts do not imply any endorsement by the author and Michigan State University. They are mentioned solely for necessary information. 15 classification of the data taken in July and August gave almost the same results. much faster. However, the manual approach was Channels 5 and 7 were found to be the most appropriate for an exact outline of the deforested areas. Heller et al. (1975) evaluated Landsat image and digital data taken during three seasons. They applied conventional photointerpretation methods and computerized classifications. The LARS (Laboratory for Application or Remote Sensing) and the PSW (Pacific Southwest Forest Experiment Station) classification systems were used. Manual interpretation accuracies of 98-99.4 percent were obtained. When computer­ ized analyses were made the LARS system gave results within 15 percent of the ground truth, whereas the PSW, with less sophisticated hardware, gave results within 25 percent. The team concluded that Landsat is a Level I land-use sensor system. 11 Dodge and Bryant (1976) reported no significant differences in forest acreage for two counties in New Hampshire when comparing estimates based on Landsat data with U.S. Forest Service figures. The differences between Landsat and Forest Service estimates of the forest resources of the two counties were -6.8 and +0.4 percent. Bedfort et al. (1977) reported that on Landsat color Level I is the most general grouping in the hierarchi­ cal categorization of land cover-use for use with remotely sensed data. It is based primarily on surface cover and is designed for use with small scale imagery, e.g., Landsat imagery (Anderson et al., 1976). 16 composites the darker tones of forestlands made them dis­ tinguishable from agricultural regions in Idaho. They also found that fall imagery was best to separate forestlands from range. Mead and Meyer (1977) used Landsat digital data taken on May 23 and July 17 to map forest cover classes in north central Minnesota, employing alternative processing systems with various pattern recognition routines. They found that the maximum likelihood routine requires less training sets than the parallelepiped routine. classification than July. May imagery gave better However, natural resource managers who evaluated the resulting maps observed the classification accuracies were not accurate enough to meet the level of information needed by forest managers. Dietrich and Lachowski (1977) reported that, by applying digital processing to Landsat CCT data, they were able to map forestlands in the Philippines and obtain an acreage estimate in four months. Hoffer et al. (1978) conducted a study covering 158 counties in four states to compare estimates of gross forest resource acreages obtained by computer-assisted analysis of Landsat data with estimates obtained by conventional proce­ dures. The results were very promising. Hardy and Agar (1978) tried to distinguish and inventory commercial or potentially commercial forestlands from non­ forestland areas in Britain, as well as estimate their acreage by using CCT data obtained in March. The computer 17 classification was found to be 84 percent accurate. Manual revision of the original classification increased the accuracy to within four percent of the manual estimate. They also noted difficulty in correctly classifying such areas as suburban wooded gardens, clear cut and/or recently planted areas on mountainous steep slopes, heathlands and orchards. Bryant et al. (1978) investigated the possibility of mapping forestlands in northern Maine using Landsat digital data taken in early July and August. One of the goals of the study was to map forests using minimum ground truth data. Initial results showed many discrepancies among softwood, hardwood and mixed forests, whereas the total forest acreage agreed quite closely (97 percent) with conventional inventory data. With more intensive use of the ground truth data the differences between computer and conventional inventory data fell within 5 percent for the forest types and 1.6 percent for the total forest acreage. However, the percent difference increased substantially for forests under 100,000 acres. Furthermore, the locational agreement between the computer classification map and the 1:15,840 scale forest type map was only about 54 percent. Chaudhery et al. (1978) investigated the application of Landsat digital data in inventorying forest classes in Bangladesh. They reported that the accomplished level of accuracy, about 70 percent, was not high enough to justify the integration of Landsat data in this type of inventory work. 18 Danjoy and Sadowski (1978) applied manual interpretation and computer-assisted techniques in classifying forest resources in the Peruvian Amazon region. They found that digital processing techniques were more accurate than manual approaches in identifying and locating forests and mapping their boundaries. Multi-element classification contributed to better illustration of the forest boundaries than the classification based on information from one element at a time. Morain and Klankamsorn (1978) discussed various uses and applications of Landsat imagery for mapping and inventory­ ing forest resources and their changes in Thailand. Diazo false color imagery using only bands 5 and 7 formed the basis for the accomplishment of the various programs. Bryant et al. (1979) reported that with computer-assisted analysis of Landsat digital data, it was possible to identify and map clear-cut areas as small as three hectares acres) in northern New Hampshire. (7.5 Also, some differences in stages of regrowth in the clear-cut areas were identifiable. Acreage measurements were found to be within about 15 per­ cent of the Forest Service figures. Townshend et al. (1979) conducted a preliminary evalua­ tion of Landsat-3 RBV (Return Beam Vidicon) data taken on August 3 for forest discrimination and mapping in an intensively dissected area in south Italy. They interpreted the imagery manually with the aid of a video density slicer. The overall correct interpretation of forestlands, with more 19 than 35 percent coverage, was about 94 percent with errors of omission and commission less than ten percent. All studies included in the preceeding literature review were unique in terms of geographical area, date, vegetative cover and condition and climatological, geomorphological and geological parameters. However, accuracy determination was not standardized to allow comparison and evaluation of studies conducted under similar conditions and using different approaches and methodology. Because of that, it is difficult to determine the utility of Landsat data as an integral part in forest management, and only inferences are possible for cases conducted under similar conditions and characteristics. Based on statements and conclusions made by the authors, only general evaluations can be made about the overall status of Landsat technology to assess various forest parameters. The results of several additional studies which were conducted to evaluate the usefulness of Landsat technology to various forestry aspects are summarized in Appendix B. CHAPTER III THE STUDY AREA A. General The study area was Barry County in southwestern Michigan (Figure 1). ing this county. There were several reasons for select­ It contains many scattered forested areas of various sizes and shapes. Detailed maps were available, showing up-to-date information about the location, distribu­ tion, composition and condition of the forest resources. Almost all the major forest cover types found in Michigan are present in the county. Both good quality color infrared photography at various scales ranging from 1:24,000 to 1:120,000 and black-and-white panchromatic photography were available. . The transportation system of the county permitted quick and inexpensive access for the collection of ground truth data. Finally, the area is located close to Michigan State University. Barry County is located approximately between the geographic coordinates, north latitude 42° 25' to 42° 47' and west longitude 89° 05' to 89° 33'. It covers an area of about 559 square miles (144,780.3 hectares). The terrain varies from level in the east to gently rolling in the west. Some areas have steep slopes but 20 THORNAPPLE1 IRVING 1 1 CARLTON I ” • ^JORDAN LAKE I WOODLAND; MIDDLEVILLE ) YANKEE I SPRINGS I RU1 AND CASTLETON I GUN HASTINGS LAKE | N A SHVILLE " N _______ ^ m HOPE ORANGEVILLE B A L T I M O R E i M A PL E j G R O V EL . I 1 ------------- P R A IRI EVI LLE / BARRY 1 --------------- 4 - --------------- JOHNSTOW N1 ASSYRIA I GULL LAKE i Figure 1. Barry County, Michigan, showing its subdivision into 16 townships and location in the state. 22 elevations are not high. 850 feet (259 meters) meters) The average elevation is about above sea level and 260 feet above the level of Lake Michigan. various sizes occur in the county. with 2,611 acres (1,057) hectares) part of the county. (79 Over 300 lakes of Gun Lake, the largest is located in the western The Thornapple River traverses the county diagonally across the northern townships. There are a number of wetlands, although well-drained soils predominate throughout the county. Beliefontaine sandy loam and Miami loam soils constitute 58 percent of the county 1928). (Deeter, The level areas are glacial sandy plains, whereas the hilly areas are glacial till plains and moraines, respectively. Agriculture land covers 198,205 acres (80,210 hectares) or 55.4 percent of the county, mostly the eastern part. Agricultural activities include general farming, livestock feeding, dairying and poultry production. Major crops are corn, wheat, oats, soybeans, and.dry beans. About 26.5 percent of the county area (93,600 acres or 37,87 8 hectares) western part. is covered by forestland, mostly in the Of these forestlands, 99.2 percent is capable of producing commercial timber. Less than 1.0 percent is incapable of producing commercial timber because of restrictions due to site conditions or administrative regula­ tions. The growing stock (excluding limbs or cull trees) totals 64.2 million cubic feet (1,797,600 cubic meters) of which 61.5 million cubic feet (1,722,000 cubic meters) are 23 . hardwood (Chase, 1970). Two-thirds of the forestland is in private ownership, whereas the state owns approximately 23,000 acres (9,303.7 hectares). Most of the state forestlands are managed primarily for recreational uses. Yankee Springs State Recreation Area is the focal point of public recreation use in the county. There are two game areas. Many lakes and the Thornapple River are also used for recreational activi­ ties. A recreation inventory (Parkins, 1978) shows that the county contains many private campgrounds, municipal parks, public school and organizational camp facilities and two state-owned parks. The transportation system of the county is good. There are many miles of federal and state highways and county and municipal roads which provide ready access to all parts of the county. The climatological conditions of the county are similar to the typical midwest climate, but Lake Michigan is a great influence. The average temperature is about 48°F (27°C); the annual precipitation is about 33 inches (84 cm) of which half falls during the growing season (May-September). Dur­ ing the winter season the total snowfall is about 40 inches (102 cm). B. The prevailing winds are south to southwest. Major Forest Cover Types in Barry County North America is divided into five forest regions: Boreal, Northern, Central, Southern and Tropical (Society 24 of American Foresters, 1954). Michigan is almost exclusively within the Northern Region, except for the very southern part which is within the Central Region. Michigan's forests are classified into six major forest cover types, 12 all of which are found in Barry County. 1. White-Red-Jack Pine Forestlands consisting of white pine strobus) , red pine (Pinus (Pinus resinosa) and jack pine (Pinus banksiana) , are sub-typed as white, red, jack pine or combination of them, where one or two of the species are predominant. This type usually occupies dry sandy so i l s , but may also be found in more moist areas. In Barry County almost all of the pine stands are plantations, of which about one-fifth are red pine. 2. Oak-Hickory White oak (Quercus alba) , northern red oak (Quercus rubra) , black oak (Quercus velutina) , bitternut hickory hickory (Carya cordiformis) , pignut (Carya glabra) and shagbark hickory ovata) are the predominant species. (Carya There are also many combinations with other oaks, hickories and associated hardwoods 12 (elm, walnut, maple, etc.). Forest cover type is defined in the Forest Terminology (Ford-Robertson, 1971) as a "descriptive term used to group stands of similar character as regards composition and development due to given ecological factors, by which they may be differentiated from other groups of stands." This type occupies well drained upland soils. It is the most common type in Barry County. Elm-Ash-Maple This type is also called "lowland hardwoods" as it is found in moist to wet soils, swamps, gullies and other poorly drained areas. It con­ sists of American elm (Ulmus americana), black ash (Fraxinus nigra) and red maple (Acer rubrum) . The three species occur in various proportions, but in Barry County, red maple is the predominant species. Maple-Beech-Cherry This is one of the forest cover types included in "northern hardwoods." It consists of sugar maple (Acer saccharum), American beech (Fagus grandifolia) and black cherry (Prunus serotina). This type generally occupies fertile, moist, well drained upland soils. In Barry County this type consists of well stocked stands. The species in this type are also the predominant understory species in other types. Aspen This is a sub-type of the aspen-birch forest cover type, consisting of bigtooth aspen (Populus grandidentata), quaking aspen (Populus tremuloides) and balsam poplar combinations. (Populus balsamifera) in various It is found on all kinds of soils except those with very high moisture content. It is a pioneer species on clear-cut or burned areas, and is not an important forest type in Barry County. Spruce-Fir This type contains white spruce (Picea glauca) , black spruce (Picea mariana) and balsam fir (Abies balsamea), is found mainly on upland loamy soils, and in Barry County it consists only of planta­ tions . Mixed Hardwoods In Michigan, and more specifically in Barry County, there are many forest areas occupied by mixtures of various hardwood species. They are simply mixed forests difficult to classify. This forest type is found on many soil types, and has various stocking levels and stand classes. Northern Conifer Swamps The principal species are tamarack (Larix laricina) and eastern hemlock (Tsuga canadensis) . The type is found on low, poorly drained soils spread throughout the southern part of Barry County. It covers only a small percentage of the total forest acreage in the county. Locust Locust is found only in planted stands in the county. These consist of black locust (Robinia 27 pseudoacacia) and honey locust {Gleditsia triacanthos) . It is not important in the county because it occupies a small acreage. CHAPTER IV DATA ACQUISITION A. Landsat Data The selection of the best available imagery is an important and fundamental step in employing Landsat data. However, what is considered as the ''optimal” imagery varies among the different disciplines which make use of this type of information. It is further influenced by the particular objectives being sought. The characterization of imagery as "optimal" depends on many factors. can be divided into two categories. For clarification these The first includes factors which are standard for all types of tasks, whereas the second includes factors which vary according to the project. The following factors are included in the first cate­ gory: 1. Good radiometric quality. Sometimes one or more of the Landsat spectral bands are poorly detected and transmitted to the receiving stations. With the exception of band 6 (which is not commonly used), the other three bands are usually of good radiometric quality. 28 2. The extent and location of clouds. A cloud-free study area is a prerequisite for nearly every investigation, especially in cases where intensive and detailed analysis of small areas is required. 3. Data from the new EDIPS (EROS Digital Image Processing System) system. Since February 1, 1979 standard processing routines (radiometric and geo­ metric corrections, haze removal and edge enhance­ ment) have been applied to the data on an operational basis. This has been done to improve the spatial and radiometric resolution of the products and increase the interpretability of various ground features. The following factors are included in the second cate- 1. The specific purposes and objectives of the study. Specific cases require special selection of images. For example, quick and accurate mapping and acre­ age estimation of burned or insect/disease-damaged forested areas require images acquired just after the event. On the other hand, the recording of changes in the forest environment occurring over time requires the acquisition and analysis of multitemporal data. 2. The time of the year. The specific features or conditions of interest are sometimes spectrally 30 more distinguishable from their surrounding areas at a specific time of the year. 3. The availability of funds. If enough money is available, there is a flexibility in acquiring complementary images and maximizing the degree of extracting the desired information. All of these factors were taken into consideration in selecting the Landsat images utilized in this study. A search was carried out of all available Landsat images of the area (i.e., nominal center of path 23 and row 30 of the Worldwide Reference System) 13 (Figure 2). Although the reference forest cover type maps were constructed from the interpretation of aerial photographs flown in 1974, the Landsat images were selected from the 1979-80 collection (after February 1, 1979). This made use of the improved quality of the new EDIPS system product. Changes in the forest resources of the county during the period between 1974 and 1979 were recorded and considered during the various calculations. Based on objectives of the study it was concluded that "optimal” imagery would be a cloudless winter scene with ^ Nominal center is the theoretical geographical tude, latitude) center of a scene. (longi­ Path is the theoretical longitudinal center line of a scene, corresponding to the center of an orbital track. Row is the theoretical latitudinal center line of a scene. 31 2,7 26 2,5 2.4 2S 29 30 Figure 2. Position of the study area within the Landsat scene of path 23 and row 30 of the Worldwide Reference System and location of the centers of the scenes required to cover the whole state. 32 snow cover on the ground. Usually there is no snow on the trees with the exception of the first two to three days following snowfall. Under this condition, although there is snow cover on the ground (especially in the leafless hard­ woods forest areas), the tonal contrast between woodland areas and their surroundings is very high. evident: The reason is part of the high spectral reflectance of the snow is obscured by the woody and branching part of the trees resulting in a lesser amount of radiation reflected from these areas and recorded by the Landsat sensor system. Two scenes which were taken on February 17 and 26, 1979 from Landsat 2 and 3, respectively, were found to fulfill the foregoing basic requirement. The second scene was eventually selected for evaluation because the tonal contrast of the forestlands with the snow covered surroundings was higher. For completeness and for comparison purposes, an additional Landsat scene, taken during the growing season, was selected. Investigation showed that the only cloud-free scene of the study area was taken on September 12, 1979, Figure 3 shows the standard false color composite of the selected fall scene with the boundaries of the study area delineated. Black-and-white positive and negative, 9 x 9 inches (225 x 225 mm) transparent imagery of all four bands and standard false color composites of both February 26 and September 12 scenes were purchased from EROS Data Center. Table 1 shows the identification number of the selected Figure 3. Barry County study area delineated on a false color composite imaged by Landsat-3, September 12, 1979. 12SEP79 C K M 3-05/H R 5-31 USGS-EDC N N 4 3 -C /M B 5 -2 1 H 5 T D SUN EL43 H139 G3H-CP-N LJ MHSR LflWBHT E -3 C B 6 -1 M M -8 35 Landsat-3 scenes, the date of acquisition and also includes * r a qualitative description of each scene. Table 1. Characteristics of the selected Landsat scenes. Identification #__________Date_________ Qualitative Description E-30358-15475-5 February 26, 1979 winter scene, snow covered, hardwoods leaf­ less , no agricultural activities, good radiometric quality, high contrast, cloud free E-30556-15460-7 September 12, 1979 fall scene, no snow, hardwoods with foliage, agricultural activities, low radiometric quality, low contrast, few clouds An additional comment on the selection of Landsat imagery can be offered. It is difficult, sometimes impossible, to find scenes which meet the specific requirements of an in­ vestigation. Occasionally researchers must compromise their needs or deviate from initial objectives or goals in order to use available Landsat data. B. Reference Data One of the most important and fundamental needs of remote sensing research is accurate reference data. The reference data required depends on the characteristics and objectives of the particular research project and on the characteristics of the remote sensor (type, radiometric resolution, spatial fidelity and spectral sensitivity). Ground information, ground truth or surface observation are 36 some other terms for reference data. However, these terms imply on-the-ground acquisition of the data which sometimes is impossible or not economically feasible. It is generally understood that information extracted from aerial photography should not be referred to as ground data. In this study the primary source of reference data was a recently completed forest cover type inventory of the county (Tatem, 1978). As stated before, the forest cover types were mapped from color infrared aerial photography at a scale of 1:31,680 flown in 1974. Four such maps were compiled, each containing four townships. The maps may be subject to some error in the interpreter's delineation of forest boundaries vis-a-vis ground conditions as all areas were not verified in the field. In cases where discrep­ ancies in delineation of the boundaries between the forest cover maps and the maps created from manual interpretation of the Landsat data (interpretation maps) were found, sup­ plementary information was used. The major type of supple­ mentary information was 1:24,000 color infrared photography flown in 1978 which were provided by the Office of Land Resource Programs in the Michigan Department of Natural Resources. CHAPTER V STUDY MATERIALS, DESIGN AND DATA COLLECTION A. Study Materials ■ Four Landsat images from two seasons (late winter and early fall) were employed in this study; all were positive film transparencies. Three of the images were standard Landsat products and the other was an enhanced image of the forest resources using the diazo process. They were: 1) a black-and-white image of band 5 (winter scene); 2) a standard false-color composite (winter scene); 3) a standard falsecolor composite (fall scene); and 4) a diazo false-color composite (fall scene). A discussion of these film products follows. 1. Standard Images: winter black-and-white; winter color; fall color. The first Landsat image to be analyzed and evaluated was a black-and-white band 5 image (1:1,000,000 scale) taken on February 26, 1979. All the available band images, both positives and negatives, for this date were subjectively evaluated to determine the best band image for forest identi­ fication. The negative images were considered inferior because the forest boundaries appeared fuzzier than on the 37 38 positive, the possibility of misclassifying water areas as forests was higher, and the interpreter lacked familiarity and experience with the appearance of tones on negative imagery. A different approach was followed for the selection of the best positive band. All four black-and-white band posi­ tive images were projected at a scale of 1:50,000. For each image a quick interpretation of the forest resources within the training townships was conducted. A subjective evalua­ tion of the four bands showed that band 5 (Figure 4) most clearly and accurately depicted forestlands because of the high spectral contrast between forestlands and surrounding snow-covered areas. The high contrast was due to the low reflectivity of the wooded areas in comparison to the high reflectivity of the snow. Because of snow and above-freezing temperatures prior to the winter image, the reflectance of the snow cover was at a maximum (O'Brien and Munis, 1975) in band 5. In addition, water areas were more clearly differ­ entiated from coniferous forests. Standard false-color composites of both late winter (Figure 5) and early fall (Figure 3) were evaluated next. A false-color composite of a scene is generated by printing three MSS bands onto a color film. The printing is typically done through blue, green and red filters for bands 4, 5 and 7, respectively. The product is referred to as "false" color because it simulates the color appearance of color infrared film. False-color composites were included in the Figure 4. Landsat black-and-white band 5 subimage of Barry County, Michigan taken on February 2 6 , 1989. Figure 5. Landsat false-color composite subimage of Barry County, Michigan taken on February 26, 1979. 41 evaluation because they are a standard Landsat product. Also, a composite, as a color multiband product, combines and complements the information which exists in each band and improves the ability to distinguish the identity and condition of various features because the eye can distinguish many more color variations than, gray shades. 2. Diazo Color Composite Although Landsat data provides a large amount of infor­ mation, it is difficult and sometimes impossible to extract all the information in a scene from standard images (blackand-white and false-color composite). Landsat data can, however, be enhanced to increase the appearance (contrast) of particular areas of interest such as forests. There are two types of contrast enhancement methods: digital, where the digital data from the Landsat computer compatible tapes (CCT's) are employed; and optical, where the photographic Landsat images are used. Only the optical enhancement method was used in this study because digital methodology was too costly. There are two major alternative optical contrastenhancement methods: cessing. photographic processing and diazo pro­ Diazo processing was selected since photographic processing is also a relatively costly procedure requiring a photographic laboratory and experienced technical per­ sonnel. The diazo process is, however, capable of producing results similar to those produced by photographic processing or digital methodology. 42 Diazo color composites are made by superimposing two, three and sometimes more diazo film copies of different spectral band images. The copies are produced by exposing Landsat black-and-white transparencies onto diazo films of different colors. The process and the equipment used in this procedure are described in Appendix C. The imagery taken on September 12, 1979 was used for the diazo processing. Positive and negative transparencies of all four bands were used in the process. These procedures were followed to determine the best diazo color composite to enhance the forest resources in the study area. Density measurements were taken with a densitometer on forest areas clearly defined on all black-and-white band positive and negative images. Five density measurements were taken on bands 4 and 5 and seven measurements were taken on bands 6 and 7. More density measurements were taken on bands 6 and 7 in order to cover the whole density range of the forest resources because coniferous species have different reflec­ tivity than hardwoods. To get reliable density measurements the selected forest targets were considerably larger than the aperture diameter of the densitometer which was one millimeter. At the Landsat image scale of 1:1,000,000 this represents an area of 200 acres (81 hectares). Most of the density measurements were taken within and directly surrounding the study area in order for the sample 43 values to be representative of the spectral signature the forests in the area. 14 of The average value of each set of these measurements was then calculated and used for the determination of the exposure time based on the sensitometric curves of the specific diazo films. Several combinations of exposure time and image type were used to produce diazo com­ posites . Each diazo composite was then evaluated in terms of interpretability by ocularly comparing the forest boundaries and patterns within the training areas on the diazo color composites and the forest cover type maps. After a thorough examination of all the diazo products and combinations, the composite chosen consisted of one yellow diazo copy of band 4, three magenta copies of band 5 and one cyan copy of band 7 {Figure 6). The corresponding exposure values were 1400, 1700 and 350 light units.15 The use of more than one diazo copy of band 5 improved the interpreta­ bility of the forest resources on the diazo composite. The color appearance of the diazo composite was similar to the standard false-color composite. The diazo film selected contained linear, regularly spaced stripes which deteriorated 14 Spectral signature is the spectral identification (characterization) of an object. It is defined by making quantitative measurements of the properties of the object at several wavelength portions of the electromagnetic spectrum. 15 The exposure values are represented to show the relative exposure relation between the various diazo films employed. However, they cannot be replicated without the use of the same exposure equipment. Figure 6. Landsat diazo false-color composite subimage of Barry County, Michigan taken on September 12, 1979. 45 the quality of the product. The direction of the linear artifacts was either parallel or perpendicular to the long side of the diazo films. The random contact of the diazo films with the Landsat images during the exposures resulted in a random positioning of the stripes on the diazo Landsat images. Thus, the stripping effect could be circumvented by producing multiple diazo copies which would shift the position of these artifacts. Two important observations were made during the inter­ pretation and storage of the diazo composites. First, al­ though the diazo films were fully developed, after approxi­ mately an hour of use, a discoloration of the projected area of the image was noticed. The degree of discoloration increased from the edge to the center of the projected area. To avoid the problem the images were projected in small portions. This was accomplished by putting a black poster board mask with a very small hole between the light source and the diazo image. The projected area was large enough to conduct the interpretation and yet the cumulative amount of light striking the entire image was kept to a minimum. A more expensive alternative would be to photograph the diazo composite with color transparent film. Second, it was also noticed that shelf storage of the diazo composites affected the film layers. Specifically, the tone variation of the yellow and cyan diazo films shifted to lighter den­ sities and the magenta film shifted to lighter densities and turned to a red color. The shifting of the tones to lighter 46 densities was probably due to ultraviolet sunlight in the storage area. It is difficult to determine what caused the magenta color to turn to red. Developed diazo films should be stored in protective envelopes, preferably within cabinets or in other dark areas. B. Study Design and Data Collection 1. Selection of the Training Set The amount of training an interpreter receives before actual manual interpretation affects the accuracy of the interpretation. It is very important for the interpreter to know how specific features or conditions on the ground appear on the various Landsat images. The interpreter's interest increases greatly when interpretation decisions are not between abstract code names or numbers but between wellknown and defined features. Basic principles of photointerpretation were used during the visual interpretation of the Landsat images. The bulk of the information was extracted from pre-specified associations between the gray tones (in the case of the blackand-white individual bands), or color renditions (in the case of the false-color composites; and ground features. The shape of the forestlands was another significant factor during the interpretation process because many of the scattered forests had easily recognized geometrical characteristics. Shape was also of great assistance during the interpretation of the winter imagery. Textural differences between forests 47 and non-forests were of greater help in the interpretation of the fall images than the winter ones. Pattern and size were of minor importance in the interpretation and mapping process, but were definitely useful in registering the overlay (called interpretation map) with the projected imagery. The first step in training was the identification, location and selection of forestlands representative of the overall forest variation in the area. That is, the woodlands sampled should represent stands of all stocking classes and of all forest cover types found in the county. Furthermore, they should be of relatively large acreage with well-defined boundaries for easy visual detection by the interpreter. Various stocking classes should be included because different percentages of ground cover reflect different amounts of solar radiation. Thus, each combination of a forest type and stocking level will be a unique spectral situation. The woodlands selected for training may be spread over the entire county or may be located within a specified area. For pro­ cedural reasons, concentrating the training set within the same area is preferred. Thus, for this study, it was decided to use Thornapple and Yankee Springs Townships as training areas. These townships contained many of the forest cover types/stocking levels/stand sizes/acreage size combina­ tions desired. The same townships were also used for the diazo processing experimentation and preliminary evaluation of the interpretability of the diazo color composites. Training on each set of data was conducted just before 48 the actual interpretation of the specific set. Thornapple and Yankee Springs Townships were excluded from further consideration during the actual interpretation, manipulation and evaluation of the Landsat system data. 2. Interpretation Procedures In this study the visual interpretation of the forested areas on the standard and diazo Landsat images was done by only one interpreter. The main reasons were: 1) the purpose of the study was to evaluate the interpretability of forest resources from Landsat imagery and not to examine the abilities of different interpreters; and 2) the limitations of time and funds. The interpretation involved a subjective evaluation and delineation of the boundaries of the forest areas with­ out considering forest type or condition. The main instru­ ment used for the photointerpretation was a back-lighted projector (Appendix D ) . The scale of the interpreted maps was 1:50,000, because projections at larger scales increased the fuzziness of the boundaries and decreased the overall quality (sharpness) of the imagery. Before interpretation, the interpreter trained himself to recognize the appearance of the forestlands on the imagery. The time spent for train­ ing on each type of imagery was unrestricted, but it was recorded for cost considerations. The time spent for the interpretation of the forest resources within each township on each type of imagery was recorded as well. A time lag of approximately a week was scheduled before the interpretation 49 of the subsequent image to reduce possible bias in the inter­ pretation decisions. The boundaries of the county and the corresponding townships were traced on acetate overlays. The Landsat images were projected onto these overlays and carefully registered. The primary features used for registration were water areas and roads. The black-and-white winter image was interpreted first. The next imagery studied was the standard false-color com­ posite taken in September. The spectral appearance of this imagery and, more specifically, the appearance of the forest areas, were absolutely different from the previous blackand-white winter scene. Because of this, bias in interpre­ tation decisions due to the experience gained from previous interpretation should be minimal. The winter false-color composite was interpreted next followed by the diazo falsecolor composite to maintain the winter-fall sequence. The final product of the visual interpretation process was four acetate overlays {interpretation maps), containing the boundaries of all the interpreted forest resources which fell within the study area. These overlays were subsequently used in the data collection and evaluation pro­ cess to determine the level of performance of visually interpreting the forest resources on the various Landsat images. 3. Verification and Error Types Making accuracy assessments in the interpretation process is highly dependent, among other factors, on the 50 verification procedure. Verification depends on the type, quality, and the quantity of the auxiliary data collected to assess Landsat classification performance. Sampling techniques are commonly used for the acquisition of data, particularly for large a r e a s , because of the lack of infor­ mation from other sources, and the high cost involved for the procurement of more detailed data. Several sampling methods have been developed or adapted to satellite tech­ nology in order to collect all the information required for a thorough assessment of a Landsat classification. Any sampling approach, however, is subject to sampling errors. Therefore, the most accurate assessment of a classification would consist of a complete enumeration of the features of interest. In this study the existence of detailed forest cover type maps permitted a complete assessment of the interpre­ tation of forest resources from Landsat images. This was achieved through a full enumeration of the errors committed during the classification process. is subject to non-sampling The process, of course, errors, but these errors are usually very small and have negligible impact. The evaluation of the interpretation was done by com­ paring the forest boundary placements on the Landsat inter­ pretation maps with those on the forest cover type maps. The comparison was made by superimposing each interpretation map on top of the forest cover type maps. Registering the two maps was a problem because different geometric projections 51 were used for the Landsat data and the forest cover type maps. The Landsat images were in a Hotine Oblique Mercator (HOM) projection, while the base map of the forest cover map was in a polyconic projection. The different projections caused the location of classification points to differ between the two maps. To compensate for this problem, the registration was done on an individual forest tract basis. Discrepancies between the forest areas depicted on each map were delineated on a third overlay, which also showed county and township boundaries. These errors were initially classified into two general classes: omission and commission. Omission errors occurred when forest areas were not classified as forests, whereas commission errors were due to the misclassification of non­ forest land cover/use classes as forest. The various omission and commission errors were then further separated into boundary or identification errors. All errors due to misinterpretation of the actual boundaries were classified as boundary errors. On the other hand, errors due to overlooking or missing individual forestlands or to the misinterpretation of individual areas of non-forest cover/use classes as forests were classified as identification errors. Identification errors are the most serious errors because individual parcels of land of a certain class are absolutely lost by assigning them to another class. Omis­ sion errors were classified based on the forest types shown on the reference forest cover maps. On the other hand, 52 commission errors were categorized from 1:24,000 scale color infrared photography taken in 1978 using a specifically developed land cover/use classification system. Table 2 shows both classification systems. The omission errors were first characterized in terms of forest type. Nine forest type categories representing the forest resources found in the county were included in the system. Subsequently, each forest type was divided into ten subcategories, each of which represented a specific stocking level and stand size {Table 3). The commission errors were divided into seven major land cover/use categories: 1) agriculture areas — lands utilized for various agricultural activities; 2) treed bog — adverse sites that support at least ten percent ground cover but of no commercial use? 3) upland and lowland brush -potentially productive lands which support trees of less than ten percent cover and brush of various species, maturity and stocking (Table 3); 4) scattered trees — lands supporting only scattered trees with less than ten percent occupancy; 5) urban lands — classified as urban but with no indication of any type of tree or brush vegetation; 6) urban-trees, a subcategory of urban — it included urban areas with consider­ able tree vegetation such as rural, suburban or resort residential areas; and 7) water-marsh — areas covered per­ manently or periodically for a significant period of time by standing water. The second category, upland and lowland brush, was 53 Table 2. Error Class Forest vegetation and land cover/use classification system developed for qualitative evaluation of the interpretation errors. Category Pine Oak-Hickory a Q V W (0 •H § Northern Hard­ woods Aspen Symbol p 0* **P 9 It . . .Ag II Aq E 0 ‘* *E 9 Conifer Swamps Spruce-Fir Q 0 ‘•-^9 S q .•.Sg Locust B 0 -- B 9 Treed Bog K q ••-Kg A B0 l II II II Areas supporting agricul­ tural crops Adverse sites supporting trees with more than 10% crown cover Areas with brush of var­ iable maturity and stock­ ing and trees of less than 10% crown cover* Upland , Lowland Brush B 1 -- B 4 Scattered Trees T Urban U Urban-Trees UT Urban areas with tree vegetation Water-Marsh W Areas permanently or periodically covered by standing water G o II 11 O •H CO CO •rt Stands ranging from re­ generation to full stocking sawtimber* II °0* * *°9 M q • . -Mg Lowland Hard­ woods Mixed Hardwoods Agriculture Description Areas supporting trees with less than 10% crown cover Urban areas without tree or other brush vegetatioi * For further explanation of classes see Table 3. 54 Table 3. Forest stand size and stocking level categories and brush (upland, lowland) ground cover classes. Stand Size A v e . Height of Stand (meters) Stocking Level Percent Crown Cover Code Regeneration <1 nonstocked 0 0 Low 10-39 1 Medium 40-69 2 High 70+ 3 Low 10-39 4 Medium 40-69 5 70+ 6 Low 10-39 7 Medium 40-69 8 70+ 9 0-14 1 15-29 2 30-59 3 60+ 4 Saplings Poletimber 1-10 10-20 High Sawtimber 20+ , High Brush 55 further divided into four subcategories based upon the per­ centage of ground covered by brush. In this study some of the land cover/use error categories were similar to the categories employed during the compilation of the forest cover type maps (Tatem, 1978). Another category was not included in the classification system because it did not represent classification error. This category consisted of individual forestlands that were correctly interpreted on the Landsat images and delineated on the interpretation maps but were not shown on the forest cover maps. These lands were omitted or overlooked in the forest cover type maps for three reasons. First, most of these lands occupied an area of less than the minimum mapping area (5 acres) of the forest cover type maps. Second, some of these stands were missed due to interpretation errors of the other investigator. The third reason were the plantations which have been established since the acquisition of the aerial photographs which were used for the compilation of the forest cover type maps. Forestlands which had been clear cut since the compilation of the forest cover type maps were also included in the interpretation maps (Figure 7). To calculate the acreage of correctly interpreted forest resources on the various Landsat images, quantifi­ cation of the various interpretation errors was carried out. For a precise calculation of the areal extent of the errors, a cell-grid was used. The density of the grid cell was 169 cells per square inch (25 cells per square centimeter). 56 % A (^■cj a.. to ^uL/-v »r 0 ___ ° C=** f = w ^§c^% 0 . eA < ^ * ? 9 ^ B fc, q P a £ ' •&] ‘x 0'*o r 03 ° _»y w ? ( ’ o ° i < I I F o r e s ts C orrectly C om m ission Figure 7. Errors Identified O verlooked F orests O m ission Errors Forest interpretation map from the color fall image of Woodland Township, Barry County, showing commission and omission errors for different forest types. See Tables 2 and 3 for explanation of codes. 57 At the map scale of 1:50,000 each cell represented 2.5 acres (1.0 hectare) on the ground. To facilitate areal measurements and reduce rounding errors to less than a quarter of a cell (0.625 acres on the ground), which is less than the nominal resolution of the Landsat system (1.1 acres), the counting and recording of the various areal sizes was done on a quarter-cell basis. That is, the actual numbers recorded on the data sheets represented quarters of cells. Recording the various errors was done separately for each type of Landsat image. Commission and omission errors were recorded separately by township on data sheets (Tables 4 and 5) . To extract the required information, several tabulations of the original data were carried out. Tables 4 and 5 are representative examples to clarify the procedures followed. To determine the absolute number of cases assigned to each classification category or subcategc :y of errors, the cases falling within each column were summed. To determine the percent of cases in each category or subcategory, the corresponding absolute value was divided by the total number of all cases in the township and multiplied by 100. For example, the percent value of the commission boundary error A (agriculture areas) shown in Table 4 was equal to 40.3, that is omission 25 qj x 100, whereas the corresponding value of the boundary error K4 (mixed hardwoods) of Table 5 was 6.8, that is ^ decimal place. x 100. All values were rounded off to one Table 4. An example of a commission error data sheet for winter color scene of Johnstown Township. Township ______Johnstown_____________ Typeof Imagery Winter Color COMMISSION ERRORS *. J C. ] 1.) 4 .4 17 14.5 3 .3 9 .1 Acr«M* ■ .7 It.1 * tfc* 14. 3 Ti v a in * l a r q w l t o 9 .4 2 5 a c re o . 1.5 T7J Table 5. An example of an omission error data sheet for winter color scene of Johnstown Township. Township Type of Imagery Johnstown Winter Color OMISSION ERRORS Boundary Id e n t if ic a t io n TotiT TofcaP te iN 9 * O v e ra ll Erft»* Acreage 20.4 5 9 .4 0. II 1 7 .5_ 1_K 14.4 14. 11.7 11. 2.9 1.1 2 .5 NmoI ut* 1 Percent 1.7 n.o 13.0 21.0 i.i 'TTf TT Ts 0 T T To “T.T 2 .5 9.0 Standard IMvIlltlM 1.1 o v e r a ll n *(» * lA o Ei ro r In th e tow nship _1i tere**n 229.0 •The u n it value la equal to 0.025 acres. 177 U1 i£> 60 In order to determine the various acreage values on the data sheets, the following procedure was used. originally recorded values First, the {number of quarter cells) were transformed to acreage values. Then the mean, the total, and the standard deviation of each category or subcategory were calculated following basic statistical procedures and formulae. The percent value of the total of each category or subcategory for the overall error assigned to each class {commission or omission) a similar approach. in the township was calculated using For example, in the case of the com­ mission errors, A, the percent value of the total acreage over the overall commission error in the township was 3 6.7, 95 0 that is q x 100, whereas the corresponding value of the omission error K 4 was 5.5, that is 12 5 q x 100. These calculated values were then used to answer various questions and to clarify, to some extent, the utility of the Landsat technology to certain forest management activities. CHAPTER VI RESULTS AND DISCUSSION The data collected from the evaluation of the inter­ pretation of the Landsat images was manipulated in various ways in order to extract the maximum information available. A number of computations, tabulations and graphs were made to: 1) illustrate characteristics and aspects of the forest resources which affected the interpretation performance; and 2) analyze the various sources of interpretation error. A. Interpretation Errors Commission and omission errors were probably due to three major factors: the sensor system and image generation process, the interpretation equipment and the interpreter. The spatial, spectral and radiometric resolution of the Landsat sensors places real limits on the accuracy that can be achieved. For example, the probability of detecting forest areas of a size close to or smaller than the spatial resolution of the Landsat system (1.1 acres) is lower than for larger forest tracts. Also, the brightness values of pixels are average reflectance values that represent the mixture of features within each pixel. Therefore, where pixels intercept the forest boundary and the orientation of 61 62 pixels with respect to the boundary line, affects boundary position and sharpness on the image. Of course, seasonal and atmospheric factors during the recording of the data are also very important. For example, snow on the ground may completely cover young forest regeneration areas and/or may affect the spectral appearance of stands of lower density. Extensive haze may also affect the spectral con­ trast on the imagery. A certain amount of boundary and interpretation errors in this study were due to the drawbacks of the system and the overall climatic conditions prevailing in the area at the time the data were collected. Projection and magnification of the images by the interpretation equip­ ment affects image geometry and clarity to some extent. Considering the above limits and the inevitability of working within them, the interpreter’s experience and background is probably the most important factor. The first step in the manupulation phase was to sum­ marize the various interpretation errors and present them in tabular form. One table was constructed for each of the four Landsat images (Tables E-l through E-4 in Appendix E ) . For each Landsat image, the corresponding table contains the frequency of boundary and identification commission and omission errors and their acreage in the form of absolute and percent values. The percent values refer to the total commission or omission error in the township. Furthermore, the Total Percent column indicates the percent contribution of the commission and omission errors made in each township 63 to the total commission or omission errors of the county. The summarized commission and omission error results of the whole county in absolute and percent values are shown in Table 6. The absolute values in the table illustrate the number of error cases and the total acreage involved in the errors arranged by type of error and Landsat image. 1. Commission Errors The smallest total commission error, in terms of number of cases and acreage occurred during the interpretation of the winter standard color composite (henceforth referred to as winter color). Commission errors are shown in Table 6. The second best imagery, in terms of commission error, was the fall diazo color composite (henceforth referred to as diazo), and third was the winter black-and-white image of band 5 (henceforth referred to as black-and-white). The poorest results, in terms of acreage, were given by the standard fall color imagery (henceforth referred to as fall color). The total commission errors of the images expressed in percentages of the total forest acreage in the study area were 5.3, 12.0, 15.5 and 16.3, respectively (Figure 8). The low commission error of the winter color imagery may be due to two factors: 1. The time of the year the imagery was taken. that time, February 26, 1979, snow 12 inches deep (Weather Bureau, U. S. Department of Agriculture) county. covered the whole The snow cover increased the spectral contrast of At Table 6. Errors of interpretation performance by type of Landsat image Barry County, Michigan. TYPE Black-and-Hhite Winter Frequency Acreage Frequency Diazo Fall Acreage Frequency Acreage t Abs. t Abs. t Abs. t 482 87.6 2603.0 83.8 1127 96.1 6957.4 96.9 6.6 68 12.4 503.8 16.2 46 3.9 222.6 3.1 9752.5 100.0 550 100.0 3106.8 100.0 1173 100.0 7180.0 100.0 85.6 7294.4 87.2 878 91.1 3796.8 89.1 936 86.0 5465.1 86.7 172 14.4 1068.4 12.8 86 8.9 465.5 10.9 152 14.0 839.0 13.3 1189 100.0 8362.8 100.0 964 100.0 4262.3 100.0 1088 "#0.0 6304.1 100.0 t Abs. t Abs. « Abs. t 1255 88.8 7988.9 86.1 1191 92.2 9113.9 93.4 Identification 15B 11.2 1293.6 13.9 95 7.8 638.6 Total 1413 100.0 9282.5 100.0 1291 100.0 591 90.8 2670.8 89.1 1017 60 9.2 326.2 10.9 651 100.0 2997.0 100.0 Boundary IMAGE Color Winter Abs. Error d H m m L A N D S A T Color Fall Acreage^ Frequency OP for Abs. u Boundary 0 ■H * m -H e Jdenti fication Total *Tolal forest acreage: 2 59,876 acres. The .icreaqe values represent the number of acres incorrectly interpreted. and graphs. They are expressed in acres in all subsequent tables Percent Interpretation 10 Error B&W Winter Com m ission Figure 8. Color Fait Error Color Winter Diazo Fall O m ission Error Interpretation commission and omission errors expressed as percentages of the total forest acreage in the study area by type of Landsat image. 66 the forested areas against the surrounding land cover/use classes. Therefore, the boundaries of many forest tracts appeared very sharp and well-defined on the imagery and, subsequently, were well depicted during the interpretation process. Furthermore, the complete or partial cover by snow of shrub lands, wetlands, brushlands and agricultural lands, resulted in an image in which very few land cover/use categories could be misinterpreted as forests. This fact reduced the possibility of making commission errors and, subsequently, improved the interpretation performance. 2. The color tone variation of the image also contri­ buted substantially to the improvement of the interpretation performance and the reduction of the commission errors. The color factor lessened the misclassification of agriculture, brushland, treed bog, water-iced and marshy areas as forests. For the black-and-white imagery, the lack of color tone and of multiband depiction was the primary factor which con­ tributed to the substantial increase of the commission errors. Portions of agricultural areas within or adjacent to forestlands and treed bogs and brushlands were often misinterpreted as forests. The misclassification was attri­ buted primarily to the similar gray tone appearance of these areas and forests on the imagery. errors were boundary errors. Most of the commission In most cases non-forest vegetation obscured the high reflectance of the underlying 67 snow to a,variant degree, causing misclassifications. The high number of boundary errors may be caused by an agricul­ tural or shrub zone of variant width along the edges of forest tracts. This zone had little or no snow on the ground at the time the imagery was taken. It was distinguishable on the winter color image but not the black-and-white. The two fall images, color and diazo, had relatively high commission errors in comparison to the winter color scene. The resultant high total commission error during the interpretation of the fall color imagery was due pri­ marily to the low radiometric quality of the scene. The spectral contrast between certain vegetation types (e.g., treed bog, brushlands) and the forest resources was low, a fact which created difficulties during the interpretation process and, in some cases, raised questions concerning decisions for the right classification. A certain degree of improvement was expected through the diazo process. The diazo imagery improved the total commission error by 2,572.5 acres (1,029.0 hectares) over the fall color imagery. In percentage values, the improvement was 4.3 of the total forest acreage in the county. In the case of the Landsat fall imagery, the low radiometric quality of the individual black-and-white Landsat images lowered the effectiveness of the diazo processing. The spectral similarity between the forest areas and other land cover/use types (Brushlands, Urban-Tress and Treed Bog) prevented a complete and absolute visual separation of the forest resources and a sharp 68 definition of forest boundaries. In the final diazo copy some land cover/use classes were recorded in tones similar to those of the forests, and so, differentiating between them was difficult. If a better quality black-and-white Landsat image was used in the production of a diazo color composite, it may significantly reduce the amount of commis­ sion error by increasing the separability of the forest resources from the other land cover/use classes. Separation of the commission error into "boundary" and "identification" revealed very clearly that the former cate­ gory contributed the most to the total error. In all types of imagery, the boundary error contributed more than 83.0 percent to the total commission error in terms of acreage, and more than 87.0 percent in terms of frequency of error cases (Table 6). The boundary error percentages of the winter images were below 90.0, whereas those of the fall images were larger (Figure 9). Black-and-white and diazo images had the highest percentages in terms of boundary error acreage. Identification error, on the other hand, was relatively small for all types of Landsat images. cation error The lowest identifi­ (Figure 9) for both acreage and frequency was accomplished with the diazo imagery. The sequence of the other images in terms of acreage was fall color, black-andwhite and winter color. In summary, the winter color imagery was superior to all other images in terms of frequency and acreage of rO 100-1 25 75 Percent Boundary Commission Error 5 q I__ --60 75 2 5- B&W Winter Color Fail N um ber of e r r o r Figure 9. Percent Identification C om m ission Error cases Color Winter Acreage 100 Diazo Fall of e r r o r cases Percent boundary and identification commission errors by type of Landsat image. 70 commission error. follows: The sequence of the other images was as diazo, black-and-white and fall color. The diazo imagery improved the commission error acreage substantially over the fall color, however, the absolute value was rela­ tively high. It was also seen that the primary source of the commission error during the interpretation of all Landsat images was boundary error, whereas the identification error contributed only marginally. 2. Omission Errors The second major category of interpretation error was omission error. The smallest total omission error in terms of number of cases and acreage was accomplished by inter­ preting the black-and-white imagery (Table 6, Figure 8). The total omission error acreage of this imagery was 5.0 percent of the total forest acreage in the county. The corresponding percentages of the other images were 7.1 for the winter color, 10.5 for the diazo and 14.0 for the fall color. Also, the frequence of the error cases increased proportionately to the omission error, in acres, in all images. The existence of snow on the ground (as in the commis­ sion error) improved the spectral contrast of the forestlands and their surroundings in the winter images. Comparison of the omission error acreages of the winter images indicated that the acreage of the winter color imagery was larger than that of the black-and-white image. These results were not expected since a color composite image should have improved 71 interpretability. Examination of the various factors con­ tributing to the creation of the errors showed that color appearance of the forest areas was the primary factor. Actually/ the spectral contrast between forested areas and the snow-covered background was higher in the black-and-white than in the color imagery. The number of omission errors of both fall images were almost double those of the blackand-white. In terms of acreage, the error of the diazo was double and that of the fall color almost triple the error of the black-and-white imagery. Thus, in terms of omission errors, the winter images were better than the fall images and the black-and-white had the fewest errors. Separation of the omission error into "boundary" and "identification" error (Table 6 and Figure 10) show almost the same results as for the commission error. The boundary error acreage was over 86.7 percent of the total omission error acreage for all images. The frequency of boundary omission error cases was over 85.6 percent of the total error for all images. On the other hand, the identification omission error was very small, less than 13 percent for all images, compared to the boundary error. In summary, the black-and-white winter imagery was the best in terms of frequency and acreage of the omission error. The winter color was the next best, followed by the diazo and fall color. Between the fall images diazo pro­ cessing improved the total omission error over the fall color. As in the case of the commission error, the boundary 100 Percent Identification Ommission Error Percent Boundary O m ission Error B&W Winter Color Fall N um ber of e r r o r c a s e e Figure 10. Color Winter Acreage Diazo Fall 100 of e r r o r c a a e s Percent boundary and identification omission errors by type of Land­ sat image. 73 error, in terms of frequency and acreage, was the primary source of omission error for all Landsat images. Identifi­ cation error contributed only marginally which implies that the possibility of missing individual forest tracts during interpretation is very low. 3. Commission and Omission Errors— Combined Effects So far the commission and omission errors have been dis­ cussed separately. However, both types of errors affect the interpretation performance in different ways. Commission- type errors add forest acreage to the actual total forest area, whereas omission errors subtract acreage from the total. Prom Table 6 it was seen that the acreage of the total commission error was higher than that of the total omission in all but the winter color image. This indicates that the acreage of the forests in the study area was overestimated in all of the Landsat images, except for the winter color imagery (Figure 11). The large overestimation of the forest resources for the black-and-white is due to the big relative difference between the commission and omission errors with the commission error being the larger. For the other three images, the differences between the two types of errors were relatively small. So far, from the separate analysis of the commission and omission errors, the winter color and the black-andwhite images were recommended as the best images, respectively. However, consideration of both types of error together (acres) 67,000-1 64 .000 - Forest Area 62 ,000 - ■ I 59 ,876 - Actual 'A creage T 58 ,000 - 57 ,000 - Figure 11. B4W Winter Color Fall Color Winter Diazo Fall Variation of the estimated total forest area (in acres) by type of Landsat image. 75 indicates that the black-and-white imagery substantially overestimated the forest resources in the study area com­ pared to the other three images. Examination of the respec­ tive errors for the other images revealed the superiority of the winter color imagery over the fall images. Thus, the winter color imagery is recommended as the best overall imagery for the interpretation of forest resources. If this imagery is unavailable, diazo processing of fall imagery is recommended. B. Accuracy Analysis So far, the discussion has dealt with the various errors that occurred during the interpretation process. How did these errors affect the overall accuracy of the visual interpretation of the forest resources? Because color infra­ red photos were often used as ground truth data, "agreement" is a more appropriate term than "accuracy." However, the meaning in terms of the evaluation process is the same so both terms will be used interchangeably. Three types of accuracy or agreement were defined and calculated in this study: 1) classification agreement which is the percent value of the size of correctly interpreted forest resources over the total reference (actual) resource size; 2) interpretation agreement which is the percent value of the size of correctly interpreted forest resources plus the area of the commission error over the total reference (actual) size of the forest resources; and 3) mapping 76 agreement which is the percent value of the size of correctly interpreted forest resources over the size of the total area displayed on the map after the evaluation. The use of a certain agreement type depends exclusively upon the established objectives, needs and requirements of the forest manager. 1. Classification Agreement Classification agreement expresses the accuracy of interpretation, taking into account only the omission classification errors. Frequently, a forest manager is interested in how accurately individual forest tracts were identified, without taking into consideration the misinter­ pretation of non-forest classes as forests (i.e., commission errors). The formulae for the calculation of classification agreement are given below; equals absolute acreage agree­ ment, K 2 equals percent agreement: K1 » T - 0 K2 = ^ x 100 where: K = classification agreement of the forest class T = total acreage of the forestlands in the county 0 = total acreage of the forestlands which were not classified as forests (omission error) The two expressions of the classification agreement, and K 2 , were calculated and tabulated for all types of Landsat imagery (Table 7). According to the calculated figures, the classification Table 7. Classification, interpretation and mapping agreement by type of Landsat image. T Y P E Agreement Black-and-white Winter OF L A N D S A T I M A G E Color Winter Color Fall Diazo Fall Acreage % Acreage % Acreage % Acreage % Classification 56879.0 95.0 51513.2 86.0 55613.7 92.8 53571.9 89.5 Interpretation 66161.5 89.5 61265.7 97.7 58720.5 98.1 60751.9 98.5 Mapping 56879.0 82.2 51513.2 74.0 55613.7 88.3 53571.9 79.9 Total Forested Acreage 59876 78 agreement for all types of imagery was over 85 percent (Figure 12). Of them, the best results (95.0 percent) were given by the black-and-white imagery, followed by the winter color (92.8 percent), the diazo imagery (89.5 percent), and the fall color imagery The forest acreage (86.0 percent). of the study area was underestimated in all cases because the value of the omission error was subtracted from the reference value (total) in the formula. In summary, the black-and-white imagery gave the best classification agreement, followed closely by the winter color imagery. On the other hand, if a 90.0 percent classi­ fication agreement is assumed to meet the requirement of a forest inventory, then the diazo imagery almost fulfills this condition. The diazo technique appears to be very promising. 2. Interpretation Agreement Interpretation agreement expresses the degree with which the forest resources were interpreted on the Landsat imagery, taking into account not only the omission errors but also the commission errors. The forest manager may be simply interested in getting only a total acreage estimate of the forest resources and not the details of the interpretation errors, the error characteristics (omission-commission) and location. If so, then the interpretation agreement provides the required in­ formation. This type of accuracy has been used in many cases in the past; however, the results do not reflect the actual 100 95 Percent C lassification 90- Agreement 85 - 80 B&W Winter Figure 12. Color Fail Color Winter Diazo Fall Percent classification agreement by type of Landsat image, 80 situation of the interpretation performance. Use of both the omission and commission errors to calculate the inter­ pretation agreement (accuracy) usually results in higher agreement values than those for classification agreement. The formulae to' calculate the interpretation agreement are given below. 1^ is the absolute agreement, I 2 is the percent agreement. Ix * T + C - 0 12 = 2 - 2 1x 100 where: I = interpretation agreement of the forestlands C = total acreage of non-forest which is inter­ preted as forest (commission error) T and O are defined as in classification agree­ ment Because of the use of both types of errors in the cal­ culations of the interpretation agreement, the forest acreage was either over- or underestimated, depending upon the absolute values of the commission and omission errors. The two errors balance each other so that the percent agreement approaches 100. Therefore, the forest manager should be cautious in using the results of this type of accuracy. If the value of the commission error is bigger than the value of the omission, the interpretation agreement given is an overestimation of the actual acreage, and vice-versa. Data from Table 6 was used to calculate absolute and percent interpretation agreement by Landsat image (Table 7). The interpretation agreement of all types of Landsat imagery 81 was higher than 89.0 percent (Figure 13). pretation agreement of the forest resources The best inter­ (98.5 percent) was for the diazo imagery, where forest acreage was over­ estimated by 875.9 acres (350.4 hectares). The winter color image had 9 8.1 percent agreement and underestimated by 1,155.5 acres acres. (462.2 hectares) the total number of forest This was the only case of interpretation agreement where the forest acreage of the study area was underestimated. The interpretation agreement of the fall color imagery was very close to the above values (97.7 percent), whereas the agreement of the black-and-white imagery was less than 90.0 percent. Thus, all three images (except the black-and-white) are acceptable when an interpretation agreement of over 90 per­ cent is desired. However, it should be emphasized again that this type of agreement (accuracy) is rather misleading. It does not display the actual affect of the various types of interpretation error on the overall classification condition. Therefore, this agreement or accuracy should be used cautiously and only if a simple number representing the accuracy of the interpretation effort is desired. On the other hand, in operational cases, it is the easiest to calculate because omission and commission errors do not have to be identified. It typically provides the highest accuracy and is, thus, frequently used in demonstrations and other activities related to the transfer of the new technology. It is also applied in extensive regional inventories of natural resources 100-, Percent Interpretation Agreement B&W Winter Figure 13. Color Fall Color Winter Diazo Fall Percent interpretation agreement by type of Landsat image. 83 where only a simple estimate value is required. 3. Mapping Agreement If the forest manager requires a value which indicates the locational or positional accuracy of the interpretation map when compared with ground conditions, the mapping agree­ ment or accuracy provides this desired information (Kalensky and Scherk, 1975). Mapping agreement is expressed by the following formulae where M^ is the absolute acreage agreement and is the per­ cent agreement. = T - 0 T - 0 M2 = T ■ + ■ c where: M = mapping agreement T, 0 and C are defined as before The absolute values of the mapping agreement were the same as those of the classification agreement so they will not be discussed again (Table 7). Only the percent values of the various images will be discussed. These values best reflect the actual overall situation of the interpretation performance and they appear to be lower than the values of the other two types of agreement. Specifically, all the values of this agreement ranged between 88.3 and 74.0 percent. The best imagery was the winter color, followed by the blackand-white, the diazo and the fall color. Winter color imagery ranked first because the absolute values of both the commission 84 and omission errors of this imagery were relatively small (Figure 14). The second place mapping accuracy of the black- and-white imagery was due, primarily, to the small value of the omission error. Among the fall images the interpretation of the diazo gave better mapping agreement than the fall color because the absolute values of both the commission and omission error of the diazo were lower. Diazo processing improved the mapping agreement by 5.5 percent over the fall color. Summarizing the conclusions of the analysis, the winter color imagery gave very good results in all types of inter­ pretation performance. native. It is recommended as the best alter­ However, it may be difficult to find a cloudless winter scene with snow on the ground and, therefore, diazo imagery is recommended as a second alternative. C. Size of Interpretation Errors The misclassification problem requires thorough investi­ gation and careful consideration. The various errors which occurred during the interpretation of the Landsat images affected the overall performance, the type and accuracy of the information selected and, finally, the assessment of the potential contribution of the Landsat system to various forest management programs. Analysis of the errors by size should give a better understanding of the affect of the various images on the interpretation performance. In fact, the size of the errors 90-1 85- Percent 80- Mapping Agreement 75- 70- Figure 14. B&W Winter Color Fall Color Winter Percent mapping agreement by type of Landsat image. Diazo Fall 86 is probably the most important factor. It is beneficial to know the percentages of the errors assigned to different size classes, because the commitment of a small number of large size errors usually is more important and raises more questions than the commitment of a large number of small size errors. The small size errors can be justified more easily as they approach the spatial resolution limits of the sensor system. The errors were classified as total error and com­ mission and omission errors subclassed into boundary and identification errors. The errors were then cross­ classified into five size classes: 0.1-5.0 acres; 5.1- 10.0 acres; 10.1-15.0 acres; 15.1-20.0 acres; and over 20.0 acres. results 1. The following is a general discussion of the (see also Appendix E ) . Commission Errors Based on the data (Table E-5) more than half of the total number of commission errors made during the interpreta­ tion of each satellite image fell within the smallest size class (0.1 to 5.0 acres). However, the frequency of making commission errors of less than five acres was higher in the winter than in the fall images. Furthermore, the average size of errors of the winter color imagery assigned to this class was the smallest among all images hectares). (3.0 acres, or 1.2 The winter color imagery had the highest among all images and the diazo the highest among the fall images. The corresponding percent acreage values of the smallest error 87 size class varied slightly between the winter color, diazo and black-and-white images (Figure 15). Examination of the percentages of the commission errors of the fall images assigned to the smallest size class indicates that the diazo imagery had the largest percent error in this class. There­ fore, diazo processing improved the interpretation of the forest resources over the fall color. Generally, the total number and acreage of errors of more than five acres in size was smaller in the winter than in the fall images. This implies that the commission errors of the fall images were more significant than those of the winter images. Also, the percentages of error cases and the respective acreage values were smaller in the winter than in the fall images. For the other size classes, as class size increased the percentages of error cases assigned to each size class decreased. In the last class (over 20.0 acres), for all images except the diazo imagery, the percent total error acreage increased. This implies that, except for the diazo imagery, a few of the commission errors were of fairly large size. Both number and acreage of the error cases of all images formed distributions which are skewed toward the smaller values. The bar graph of Figure 15 illustrates very clearly the skewness of the distributions. The figure also shows a second small peak in the last class but this is due to the way the errors were grouped as this class had a much wider class-width. Most of the commission error of the winter color imagery occurred in the smallest size class. 50-i | | B4W Winter C o lo r Fall Color Winter D i a z o Fall 30-1 Percent Total Commission Error 10- 0.1 - 5.0 5.1 - 10.0 10.1 Size C lass Figure 15. - 15.0 15.0 - 20.0 over 20.0 (acres) Distribution of total commission error by size class (acres) and type of Landsat image. 89 However, the diazo imagery was slightly better than the winter color imagery when the two smallest classes were con­ sidered together. This implies that the commission errors made during the interpretation of the winter color and diazo images tend to be of smaller size than the corres­ ponding errors of the other images. Because of this, it is concluded that the use of winter color or diazo images is preferable in mapping the forest resources. The boundary and identification commission errors were also grouped into the same five acreage size classes. Sum­ mary tables were constructed and included in Appendix E (Tables E-6 and E-7). Bar graphs of the percentages of the acreage values grouped by size class and Landsat image were created for both types of commission error (Figures 16 and 17). As in the case of the total commission error, over 50.0 percent of the number of boundary errors made during the interpretation of each image were assigned to the smallest size class. The highest percentage was that of the winter color followed by the black-and-white, diazo and fall color. In terms of acreage, the winter color had the largest percent error, followed by the diazo, black-andwhite and fall color. For identification commission error, as in the case of boundary commission errors, over 50.0 percent of the number of the identification errors had a size of less than 5.0 acres. Figure 17 illustrates the distribution of the acreage of the errors by size class. It is seen that the larger part B&W W i n t e r □ 1111 C o lo r Fall Color Winter D ia z o Fall 30 Percent Boundary Commission Error ak-ltickory, M - Northern Hardwoods, A = Aspen, E = lowland Hardwoods, K = Mixed Hardwoods, Q = Conifer Swamps, S - Spruce-Fir and B ~ locust. See Table E-ll for explanation of codes. 2Thi> values in the columns under "»■ indicate the percentages of the actual acreage al the classes in the study area which were omitted durinq the interpretation process of the Landsat images. Table E-15. (cont'd.) T Y P E Forest Types/ Stand Size Black-and-white Winter Color Fall *2 Abs. I M A G E Color Winter Acreaqe % Abs. Diazo Fall Acreaqe % Abs. Reference Data Acreaqe 6 470.1 11.8 914.6 22.9 677.3 17.0 569.0 14. 3 3984 406.2 6.9 1026.0 17.5 510.7 B. 7 779.3 13.3 5853 Ks 21.9 1.4 68.6 10.5 53.8 8.3 69.5 10.7 652 o 70 K 1 *r L A N D S A T Acreaqe Acreaqe Abs. OP 19.3 8.0 60.0 24.8 21.3 8.8 77. 3 31.9 242 Op 46.3 4.8 186.8 19.5 41.2 4.3 113.1 11.8 959 8.8 12.4 8.8 12.4 0 0 9.6 1j. 5 71 0 0 0 0 0 42 2.5 8.9 0 0 28 KP °s sk SP ss nB "p Bs Tota 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2.5 4.5 0 0 56 9. 1 91.8 59.4 13. 1 8. 1 23.7 15.0 158 0 16.2 44.1 6.8 18.4 37 14.4 0 2997.0 0 10.0 B162.8 35.7 0 4262.3 0 6304.1 59876 f = Pine, O - Oak-llickory, M = Northern Hardwoods, A = Aspen, R = lowland llarduoods, K = Mixed Hardwoods, (J - Conifer Swamps, S ~ Spruce-Fir and B = locust. See Table E-13 for explanation of codes. 2Th« values in tlic columns under indicate the pcrcenl.aqes of the actual acreaqe of the classes in the study area which weie omitted durinq the interpretation process of the Landsat imaqes.