3—1: ‘ 543% ‘3 z ; guru. 1 .11... , 3m“. ‘ $5.? . ‘ 5.... Hi 2 air. . at! In . finfivflfl u .. 51‘ in.“ "a . 2.! 2.7 .. ‘ ‘ .‘ku .1I31-,I' »\.1. . ..J.|i. .. 4 .‘Lilvltllilzlul’! 1.! vt.lv..\ :0xl0rlflu . punt 01.07:. VV‘V ‘D5ulltlle\ I V. .‘w: $13; ~ . ,. . ., 1 v. . 31:1 . ‘ . , ‘IX t. . 9. ifi,..§ _ . ; éfifimknén .. k “3‘ 1-. . v F. a}? ..fi...c..h . . M1“: . ull. - i :n , w .e . ,.... .‘ . , , . . O‘ ‘ .v'- A. 3!. ~L ‘1 I J ‘ » it). » . . . .. ., , 3. .q-_‘..m.n.u..... . :.uh..uw.u.e. sis , . . V 3 ... . ‘ . ‘ ‘ .5151. .5 ‘ .. .r: THsszs fl P) / 'fA‘f\ (7’ L4 ' lHIWHIHI|||l|||l|||llllllllllllllllllllllllllllllllllllll Llfifififiy 3129302074102 l‘mchecn State Univer ity This is to certify that the thesis entitled DETECTION AND MEASUREMENT OF AMAZON TROPICAL FOREST LOGGING USING REMOTE SENSING DATA presented by Deborah Jean Janeczek has been accepted towards fulfillment of the requirements for MA degree in Geography [“4 v; /; ~— Major professor Date 12‘16’99 0-7639 MS U is an Affirmative Action/Equal Opportunity Institution '4. PLACE IN RETURN BOX to remove this checkout from your record. To AVOID FINE return on or before date due. MAY BE RECALLED with earlier due date if requested. DATE DUE DATE DUE DATE DUE SEP 1 7 10172 In: § JULY} 2.573903 .? moo c/CIRCJDCQOmpSS-p.“ DETECTION AND MEASUREMENT OF AMAZON TROPICAL FOREST LOGGING USING REMOTE SENSING DATA BY DEBORAH JEAN JANECZEK A THESIS Submitted to Michigan State University In partial fulfillment of the requirements for the degree of MASTER OF ARTS DEPARTMENT OF GEOGRAPHY 1999 ABSTRACT DETECTION AND MEASUREMENT OF AMAZON TROPICAL FOREST LOGGING USING REMOTE SENSING DATA By Deborah Jean Janeczek Because forests are complex, globally distributed ecosystems, remote sensing provides a valuable means for monitoring them. Satellite data have been used to determine the rate of deforestation in Brazil’s Legal Amazon. The majority of deforestation measured thus far has been has been done by Clear cutting, burning for pasture, and subsistence farming. An apparently new phenomenon occurring in Brazil’s tropical forests is selective logging; generally, selective logging can be detected with Landsat TM data, although it is sometimes camouflaged by the crowns of the residual trees and can be misclassified as undisturbed forest in most classification techniques. A 1988 estimate for deforestation reported by Skole and Tucker (1993) and subsequent analyses by researchers at Michigan State University and lnstitudo national de Pesquisas Espaciais do not include selectively logged areas. The purpose of this study is to quantify the area Of selective logging missed in previous deforestation estimates. It is the first basin-wide study, finding 5,309 km2 of selective logging in the 1992 Landsat TM images. ACKNOWLEDGEMENTS I would like to acknowledge the help and guidance of my committee. I extend Special appreciation to Dr. David L. Skole for the motivation and the opportunity to excel that he provided as my advisor. I would like to thank my thesis committee members, Dr. David Lusch and Dr. Jiaguo Qi for their academic expertise. Special thanks are given to William Salas, Walter Chomentowski, Mark Cochrane, and my fellow graduate students at BSRSI for their advice and support. I would also like to recognize the help provided by my colleagues Amy Sayers for her mapping Skills, and Diane Cox for her enthusiasm and support when I needed it the most. A very special thanks to Christopher Barber without his technical expertise, encouragement, and support this wouldn’t have been possible. iii TABLE OF CONTENTS LIST OF TABLES LIST OF FIGURES CHAPTER 1. THE PROBLEM Background Cryptic Deforestation Selective Logging and Fire Selective Logging and Carbon Previous Studies EXPERIMENTAL EFFORTS TO DEVELOP AUTOMATED METHODS FOR DETECTION OF LOGGING AND DEGRADATION Introduction Standard Deviation Focal Analysis Slope Analysis NDVI Texture Analysis Conclusion METHODS EMPLOYED FOR MAPPING LOGGING AND DEGRADATION Rationale Data Set Data Preparation Automated Detection of Logging Patios Digitizing Maps RESULTS Analysis of Total Logging Detected in 1992 Landsat TM Images Analysis of Canopy Degradation Digitized in 1992 Landsat TM Images Analysis of Logging and Canopy Degradation in 1992 Landsat TM Images Comparison of 1991 and 1997 Landsat TM Images Analysis Of Basin Wide Verification Analysis Of Accuracy Assessment of Logging Patios iv vi vii PAGE \IOSO'IOON-A 14 14 16 18 19 2O 21 23 23 24 25 26 29 32 32 33 34 35 36 37 DISCUSSION Automated Detection of Logging Patios Legal Amazon Basin Digitizing Maps Considerations About the Accuracy of the Method CONCLUSION 39 45 46 47 48 LIST OF TABLES TABLE Table 3.1 . x-y raster coordinates for each polygon that was digitized around areas of selective logging. x-y coordinates for row A is Obvious logging polygons. x-y coordinates for row B is subtle logging polygons. Table 4.1 Selective logging estimates for each method used in this study. Also, estimates of selective logging that was captured as agriculture in the 1992 deforestation estimate. Table 5.1 Sensitivity analysis calculating a possible lower bound estimate for Nepstad et al (1999) data and upper bound estimate for the data in this study. Table 5.2 Calculation of a lower bound estimate of the Nepstad et al (1999) data. vi PAGE 73 97 41 42 LIST OF FIGURES FIGURE PAGE Figure 1.1 Map of Brazil’ 3 Legal Amazon. This study area is an administrative area within the country of Brazil that includes 5x106 km’ of the nine states of Acre, Amapa, Amazonas, Para, Rondonia, Roraima, plus parts of Mato Grosso, Maranhao, and Tocantins. 51 Figure 1.2 Landsat TM color composite image of Para, Brazil, for path 222 and row 63, acquired on 24 July 1991. Areas oft ropical forest, deforestation, regrowth, logging patios, and canopy degradation. 53 Figure 1.3 This image is a Close-up view of figure 1.2, showing an example of cryptic logging and obvious canopy degradation. 55 Figure 3.1 Landsat TM color composite image of southern Para state, Brazil, for path 223 and row 65 acquired on 6 June 1993. A selective logging area of Mahogany that Christopher Uhl (199) used as a study site. 57 Figure 3.2 The thirty images that contain visible selective logging and were used in the logging analysis. These images are highlighted in gray and found in a crescent along the eastern and southern Legal Amazon. 59 Figure 3.3 Band five of a Landsat TM image in the state of Mato Grosso, Brazil, for path 227 and row 69, acquired on 5 July 1992. All land cover but forest has been masked out Of the image. 61 Figure 3.4 An example of logging patios that were detected using the texture analysis on band five of a Landsat TM image. Mato Grosso, Brazil, 19 May 1992. 63 Figure 3.5 Landsat TM color composite image of Mato Grosso state, Brazil, for path 226 and row 69 acquired on 19 May 1992. This image is the same image as in figure 3.5 showing the visible logging patios that were detected using the automated method. 65 Figure 3.6 Landsat TM color composite image of Mato Grosso state, Brazil, for path 226 and row 69 acquired on 19 May 1992. Logging patios that have been detected with the automated vii method and buffered using the 180m buffer specified by Souza and Barreto (1999). Figure 3.7 Landsat TM color composite image of Mato Grosso state, Brazil, for path 226 and row 69 acquired on 19 May 1992. Areas of obvious selective logging that have been digitized. Figure 3.8 Landsat TM color composite image of Para state, Brazil, for path 222 and row 62 acquired on 24 July 1991. Areas of subtle selective logging that have been digitized. Figure 4.1 Obvious logging for each scene (units km’). Figure 4.2 Subtle logging for each scene (units km’). Figure 4.3 Total Canopy Degradation (units km’). Obvious logging and subtle logging combined. Figure 4.4 Obvious logging and subtle logging for each scene (units km’). Figure 4.5 Total deforestation due to selective logging not included in the 1992 estimate of deforestation (units km’). Figure 4.6 Cryptic deforestation captured compared to canopy degradation captured (units km“). Figure 4.7 Images in study divided into four regions by extent Of logging for the accuracy assessment of the detection of logging patios. viii 67 69 71 89 90 91 92 93 94 96 CHAPTER 1 THE PROBLEM Selective logging in Brazil’s Legal Amazon is a relatively new phenomenon not included in current deforestation estimates because it iS not detected in unsupervised classifications. Recently, studies have been done on this phenomenon; however, they have not yet quantified selective logging in the entire Amazon basin nor has an automated method been developed that detects selective logging. The purpose of this study is to develop a method to detect selective logging and quantify the area affected in the Brazil’s Legal Amazon. BACKGROUND AS the largest contiguous tropical forest on the planet, Brazil’s Legal Amazon, has received worldwide attention in natural resource and environmental studies. The Legal Amazon is an administrative area within the country of Brazil that includes 5x106 km2 of the nine states of Acre, Amapa, Amazonas, Para, Rondonia, Roraima, plus parts of Mato Grosso, Maranhao, and Tocantins (Figure 1.1). The most important concern for the Legal Amazon is its dramatic speed of deforestation. According to a group of NASA-funded scientists, 6% of the primary forest in Legal Amazon had been cut down by 1988; 90% of this was cut down after the 19705 (Skole and Tucker 1993). The rapid and extensive clearing of Amazonia forest is highly correlated with the beginning of a government colonization program started in 1968, when the Brazilian federal government decided to “bring men without land to the land without men” to exploit this remote resource in the Amazon and to promote Brazil’s economic growth. Road construction and improvement is an important component of the colonization program (Dale et al., 1993), and are also considered proximate causes of deforestation in the Amazon (Pfaff 1997, Mertens 1997). With the extension of the road network, colonists moved into the frontier of the Amazon and cut the forests adjacent to the roads with the intent of establishing agricultural settlements. For more than three centuries, logging was restricted to the floodplain forest bordering Amazonia’s major rivers (Rankin, 1985). But the construction of strategic access roads into Amazonia coupled with the depletion of hardwood stocks in the south of Brazil have transformed Amazon logging from a minor activity to a major growth industry (Verissimo et al., 1992). Selective logging has several environmental impacts, including increased fire susceptibility (Holdsworth and Uhl, 1997), damage to nearby trees and soils (Johns et al 1996), increased risk of local species extirpation (Martini et al 1994), and increased carbon emissions (Houghton 1995). Furthermore, uncontrolled exploration by loggers catalyzes deforestation by opening roads into unoccupied government lands and protected areas that are subsequently colonized by ranchers and farmers (Verissimo et al 1995). ngtic Deforestation Selective logging is the process of removing four to twenty trees per hectare. Although this may seem to be a sustainable way of logging forests, it is not. Selective logging in the Amazon is not planned to minimize effects to surrounding forest. First, bulldozers are used to open a network of logging roads. Second, patches of forest are cleared at intervals along these logging roads to serve as log landings; log landings or patios are areas of forest that are cleared to stack logs waiting to be transported by trucks. Third, trees are felled and bucked. Fourth, logs are linked by Choker cable to a bulldozer or skidder by maneuvering in the bole zone. Finally, the logs are skidded to patios in preparation for transport out of the logging area (Johns et at, 1996). This process can devastate surrounding forest. In many instances, the logging roads that are created by bulldozers are not used to skid the logs to the patios; instead the bulldozer creates new roads to the patios, destroying even more forest. Also, vines in the forest canopy connect trees; thus when one tree is felled, it can potentially take down five to ten trees with it. Selective logging destroys surrounding forest and Should be considered a form of deforestation. Although visible in Landsat TM imagery, selective logging is not easily detected using an unsupervised classification and is not, therefore, included in current deforestation estimates (Stone and Lefebvre, 1998). Because logging is not readily detected using most image Classification techniques, some researchers have called it a “cryptic” form of deforestation (Nepstad et al, 1999). AS such, the area affected by selective logging in the Amazon has not yet been quantified on a basin-wide level. The process Of deforestation by logging is complex, and results in a heavily degraded forest environment. Selective logging leaves behind primary and secondary roads, patios or truck loading areas, a mixture Of Intact forest with treefall gaps, and damaged trees. Loggers do not clear-cut the forest and then burn it. Logging does not usually kill all trees but it severely damages forests (Nepstad et al., 1999). Logging companies in Amazonia kill or damage 10-40% of the living biomass of the forest during the harvest process (Nepstad et al., 1999, Uhl et al., 1991, Verissimo et al., 1992, 1995). There is little quantitative information on this new trend despite its potentially large impact in terms of carbon release, forest biomass, hydrology, sustainable development, and biotic diversity (Stone and Lefebvre 1998). Most selective logging can be visually identified in Landsat TM data by its pattern, or texture, in the forest canopy. Visual cues for selective logging in Landsat TM data include degradation in the forest canopy and increased shadow, extensive logging patios, and occurrence of primary and secondary roads (Figure 1.2). In some areas of logging, degradation in the forest canopy is not visible, but the secondary roads and extensive logging patios are, indicating logging activity. In this study, I’ll refer to these areas as cryptic logging areas (Figure 1.3). Also, for the purpose of this study, the term “canopy degradation” refers to visible disturbance in the forest canopy around logging patios indicating an area of extensive logging and possibly has been burned. Selective Logging and Fire Uncontrolled fires are an underestimated and underreported disturbance in the Amazon basin (Cochrane, 1998). Undisturbed tropical forest is generally immune to fire, while selectively logged forests are susceptible to fire. Except for tree-fall gaps and areas of unusual fuel structure, in an undisturbed tropical forest, fire will spread as a thin, slowly creeping ribbon of flames a few tens of centimeters in height (Cochrane, 1998). Damage from logging, however, can increase the fuel availability by adding debris to the forest floor, and devastating fires can result. The effect of logging is to increase forest flammability by reducing forest leaf coverage by 14-50%, allowing sunlight to penetrate to the forest floor, where it dries out the organic debris created by the logging (Nepstad et al., 1999). Although loggers often only extract four to eight trees per hectare, the resulting forest is fragmented into a mosaic of gaps and forest patches, where canopy cover is reduced by half (Uhl and Viera 1989). Post-logging fuel loads in logged forest are three times higher than in uncut primary forest, and large gaps from logging can burn after only five to six rainless days in the dry season (Uhl and Kauffman 1990). Most fires in Amazon are set intentionally in pastures and fields and then spread into nearby selectively logged areas. Both logging and fire increase forest vulnerability to future burning and release forest carbon stocks into the atmosphere, potentially doubling net carbon emissions from regional land-use during severe El Nino episodes (Nepstad et al., 1999). The average rate and intensity of forest burning and deforestation can be expected to increase as previously burned forest area expands (Cochrane et al., 1999). Selective Logging and Carbon Deforestation rates In Amazonia are used to determine human effects on the global carbon cycle. Most carbon studies include only outright deforestation but do not include logging but only accounts for clear-cut forests because forest conversion to agriculture is monitored from space easily using Landsat TM images allowing, large-scale deforestation maps to be developed. The forest openings created by logging and accidental surface fires are visible in Landsat TM images, but they are covered over by regrowing vegetation within one to five years, and are easily missed in the absence of accompanying field data (Nepstad et al., 1999). This forest impoverishment can cause a significant release of carbon into the atmosphere, which is not included in existing estimates of the Amazonian carbon balance. Nepstad et al (1999), estimated that in 1996 logging released approximately 4-7°/o of the net annual carbon release estimated for deforestation in Brazilian Amazonia. Some Of the studies Of carbon release may be underestimating carbon loss to the atmosphere due to this new phenomenon. Previous Studies Skole and Tucker (1993) used Landsat satellite data for the Brazilian Amazon Basin, in 1978 and 1988 to measure deforestation and forest fragmentation. For 1988 they used black and white photographic images using Channel five of the Landsat Thematic Mapper (TM) and then digitized the deforested areas using visual interpretation and standard vector GIS techniques. To determine the edge effects from fragmentation they extracted forest fragments which were <100 km2 and computed edge effects for a buffer zone of 500 m and 1,000 m. They found that deforestation increased significantly between 1978 and 1988 (78,000 to 230,000 kmz) as did the total affected habitat (208,000 to 588,000 km’). This study was the first widely published estimate of deforestation using satellite data analysis, but did not explicitly consider logging and forest degradation. Stone and Lefebvre (1998) examined 1991, 1988, and 1986 Landsat satellite imagery to determine the extent of selective logging in areas west and northeast of the urban center Of Paragominas, Para. They studied how fast selective logging was occurring and how long logging remains visible in Landsat TM data. All images were Classified using either a supervised or unsupervised Classification methods. They state “it is doubtful that an automatic classification procedure could be developed to define the location and extent of selective logging.” They therefore relied on a visual interpretation to define the location and extent of forests affected by selective logging. These areas were digitized manually on a computer screen. The digitized polygons of selective logging were overlaid on TM images to define how much selective logging in areas classified as intact tropical moist forest. When comparing polygons Of selective logging from 1986 and 1991 data in the western region they found no spatial overlap i.e., those areas, which were selectively logged in 1986 were not selectively logged in 1991. Also, there were no apparent visual cues in the 1991 imagery, which allowed location of the areas that were selectively logged in 1986. Of areas selectively logged in 1986, 91% were Classified as forest in the 1991 imagery and only 9% were classified as fields, pasture, and regrowth. Stone and Lefebvre also tested a texture analysis on TM band 4 (076-09 microns) to investigate whether forest canopy texture was Significantly different in logged forest from that of undisturbed forest as well as a normalized difference vegetation index (NDVI) analysis was also computed on the images. They found that texture and NDVI images were not helpful in defining selectively logged forest. The tendency was for logged forest to resemble secondary growth. Nepstad et al (1999) report that estimates of annual deforestation for Brazilian Amazonia, where one-third Of the world's tropical forests are found, capture only 60% of the total forest area that is impoverished by humans each year. The remaining 40%, they Claim, is due to logging. They state that “binary approaches such as monitoring deforestation by using imagery from Landsat TM neglects those forest alterations that reduce tree cover but do not eliminate it, such as selective logging and surface fires in standing forests.” They did not use satellite data to estimate the amount of selective logging. Nepstad et al (1999) estimated the area of Brazilian Amazonia forest that is impoverished each year through logging by interviewing 1,393 wood mill operators, representing more that half Of the mills located in 75 logging centers. They Obtained each mill’s harvest records of roundwood (tree trunks) for 1996 and 1997 and the roundwood harvest rate (m3 of timber per ha of forest). Using this information, they calculating the forest area required to supply each center’s timber production. They also estimated the area subjected to surface fire each year by interviewing 202 landholders in five regions along a 2,200 km transect through the states of Para, Mato Grosso, Rondonia, and Acre. They had the landholders draw onto satellite images the forest areas on their property that had been deforested and the forest areas that had burned by surface fire (without prior forest felling) in 1994 and 1995. Nepstad et al (1999) estimate that 10,000-15,000 km2 of undisturbed forest are logged each year (1996 and 1997) by 2,300 sawmills Operating in Brazilian Amazonia. According to Nepstad et al (1999), selective logging in 1996 affected an area of forest that was three-fourths as large as the satellite-based estimate of annual deforestation from 1992-1994 and equal to the estimate for 1991 and 1992. Nepstad et al (1999) suggests that cryptic forest impoverishment through selective logging causes a significant release of carbon into the atmosphere that is not included in existing estimates of the Amazonian carbon balance. As a result, the Brazilian contribution to the increase in atmospheric C02 has been underestimated and its success in curbing the rate of Amazonia forest impoverishment has been overestimated. Nepstad et al (1999) found that within properties surveyed for the fire study, the area of standing forest that was affected by surface fire in 1994 and 1995 (310 km’) was 1.5 times greater than the area that was deforested in those years (210 kmz). They state, “Although extrapolation of this data set to the entire Amazon is not warranted, these data indicate that the area Of Amazon forest affected by surface fire each year may be similar in scale to the area affected by deforestation.” In a detailed study of logging in a 32,520 hectare area near Paragominas, Souza and Barreto (1999) present a method for estimating selective logging. They used Landsat TM images (bands 1-5, and 7) of their study area (path222/row62) for June 1984, July 1991, and July 1996. They used a linear mixture model to identify spectrally pure pixels and estimate the soil, vegetation, and shade fractions within each pixel of their TM images. Soil exposure enabled explicit detection of logging patios. 10 After identifying logging patios, they used a buffer routine to estimate the total area affected by selective logging. During field calibration trials, they determined the buffer Size from data collected in 82.5 hectares of unplanned logged forest. The total area logged was divided by the number of log landings (n = 10) to estimate the average area of forest within reach Of a logging patio (8.25 ha). An average extraction radius Of 162 m was calculated using these data. However, it was necessary to use a multiple of the Landsat TM pixel Size (30 m) for the buffer routine, so a radius Of 180 m was used in the final analysis. They found 2,089 ha in 1984; 2,585 ha in 1991; and 662 ha in 1996 of selective logging. To assess the accuracy of estimates of the area affected by logging Souza and Barreto (1999) applied the methodology to the 82.5 hectare area of typical logged forest. Using the estimated 180 m radius Of extraction and the number of log landings in the logged area they calculated that logging affected an area Of 80.5 hectares, 97% Of the actual area. The area identified as potentially being logged included 294 of the 326 (90%) trees actually extracted from the site. In a recent and important study of fire degradation, Cochrane et al (1999) studied forest fire dynamics in the Amazon to understand the effects of this disturbance force. They did field studies in the Tailandia region using ten 0.5-ha plots (eight fire-affected and two control) distributed over 100 kmz. These Sites were established in 1996 to study fire impacts on forest structure, biomass, and 11 Species composition. After the dry season of 1997, fire recurrence, tree mortality, and biomass combustion levels within forests of different burn histories were quantified. Cochrane et al also examined characteristics of fires while they were occurring in four forest types (previously unburned, once-burned, twice-burned, and more than two previous burns) in December 1997. For each fire Observed, they measured flame heights and depths. The time the fireline took to move across a known distance was used to calculate the rate Of spread. Cochrane et al found that the first fire to enter a forest usually moves slowly along the ground and consumes little besides the dry leaf litter. In these first fires, 95% of trees >1cm dbh are killed because of their characteristically thin bark. Second fires are faster moving and much more intense because of increased flame depth. They found that large trees have little survival advantage against the second fire during these more intense fires. Cochrane et al used satellite images from Landsat TM to conduct multitemporal analyses of fire in the Tailandia and Paragominas regions. They used a linear mixture model to separate forest from nonforest and to classify burned forests in all images, they then cross-tabulated these images, which provided a history of deforestation and forest burning throughout the study regions. They found that areas that are minimally forested because of the recurrence Of fire are likely to appear deforested in satellite imagery analyses. 12 They also conducted a detailed study of deforestation in burned forests, using imagery of Paragominas for the period from 1993 to 1995 to test whether the deforestation that had burned in 1992 was intentional or accidentally induced by fire. In the Paragominas region, they estimated that accidental fire-induced deforestation increased deforestation estimates by 129% between 1993 and 1995. Cochrane et al state, “This surprising result implies that the basin-wide jump in estimated deforestation rates may have occurred largely because of the widespread forest fires of 1992 and 1993.” The purpose of this analysis was to develop a method, which would be used to detect selectively logged areas. This study had two steps: (1) developing a model that can be run on satellite images to detect cryptic deforestation; and (2) making a visual inspection of Landsat TM 1992 images and identifying selectively logged areas. The visual inspection identifies areas of canopy degradation around logging patios and these areas are then digitized. The primary objective of this analysis was to quantify the area in Brazil’s Legal Amazon that was selectively logged in 1992 and add this estimate to the 1992 deforestation data for a total area deforested. l3 Chapter 2 Experimental Efforts to Develop Automated Methods for Detection of Logging and Degradation Introduction The purpose of this analysis was to develop a procedure to define the location and extent of cryptic logging and degradation in the Brazilian Amazon forest, in order to enhance previous and current deforestation estimates. The area of selective logging targeted for detection was the portion of the forest canopy where trees have been removed but the area has not been clear-cut. The selective logging process removes 4-20 trees per hectare and leaves behind a mixture of intact forest with treefall gaps, primary and secondary roads, and logging patios. The first step before developing a method was to test the unsupervised Classification technique used on Landsat TM images to derive the 1992 deforestation data. I tested the unsupervised Classification on a Landsat TM scene that contained selective logging with 45, 50, 55, and 60 classes; in all cases cryptic logging and canopy degradation was clumped with forest. I was unsuccessful at defining a separate spectral Class for selectively logged forest. l4 Verifying the findings of Stone and Lefebvre (1998). This is because selectively logged sites are composed of a combination of intact forest canopy, damaged canopy, secondary growth forest, understorey vegetation, and bare soil, all of which are spectrally Similar to other Classes in an unsupervised classification. In most cases the logging patios and roads were also clumped in with forest. Image classification procedures group pixels into classes or categories based upon distinctive, multispectral patterns of digital numbers (ON) and are normally categorized as either supervised or unsupervised. A supervised Classifier uses training data input by an analyst that are based upon prior knowledge of land- cover at selective locations. The unsupervised classification determines the classes by spectral distinctions inherent in the data. The 1992 deforestation GIS layer was generated from an unsupervised Classification. This GIS has seven classes: forest, nonforest, cerrado, secondary growth, water, cloud, and Cloud shadow. Unsupervised classifiers do not depend on the input of a training data set. They generate classes based upon Clustering the multispectral values into groups based upon similarity (Lillesand and Kiefer, 1994). Once the Spatial clusters are generated, an analyst attempts to determine the nature Of the Clusters and provide labels. 15 All of the attempts I made to detect cryptic deforestation using automated methods were unsuccessful; however, I was successful at detecting the logging patios and incorporated this material into my study with a buffer routine (discussed in Chapter 3). I’ve listed below the methods attempted and a description of how they worked. Standard Deviation Focal Analysis In the visible part of the electromagnetic Spectrum, the Spectral pattern Of vegetation is dominated by absorption in the blue (450-520 nm) and red (630-690 nm) bands and reflectance in the green (520-600 nm) band. The dominant pigments in plants are chlorophyll A and B, and light absorption is required by plants to support photosynthesis. In this analysis, I used TM band 3, the red band, of the visible Spectrum region because Of the absorption by vegetation. Cryptic deforestation leaves behind an intact forest canopy, treefall gaps, damaged trees, patios, and logging roads. Forest leaf coverage is reduced allowing the sun to penetrate to the forest floor to dry out debris and allow secondary vegetation to grow. I hypothesized that because of these Changes, the digital numbers (DN) of the pixels in the red band would have high variation in areas of cryptic deforestation. Areas of intact forest canopy would have low DNS because of high absorption, while areas of treefall gaps would have higher DNS because Of lower absorption and higher reflectance. Also, damaged and dying trees would have higher DNS than the intact forest canopy. Logging patios 16 and roads would have the highest DNS in areas of cryptic deforestation due to little or no vegetation. Using the above information, the first automated model I tried was a standard deviation focal analysis to detect variation in areas of cryptic deforestation. Using subset, an ERDAS IMAGINE command which breaks out a portion of a large file into one or more smaller files, I extracted TM band 3 from the full data set. A standard deviation focal analysis was conducted on the resulting image using a 3x3-moving window. The focal standard deviation module returns the standard deviation of the pixel DNS in the focal window around each pixel of the image. The resulting image was re—scaled and then inserted as TM band 3 into the spectral subsets 4, 3, 2 of the image. Then an unsupervised classification was run on the image using 45 classes and 12 iterations. I then analyzed and labeled the spectral Clusters and found that the resulting Classification captured logging patios and roads that were not identified in the 1992 deforestation estimates as deforestation, but classified the areas around the logging roads and patios as forest. Therefore, cryptic deforestation was not detected. This attempt was unsuccessful in detecting cryptic deforestation because variation in the DNS of the disturbed forest canopy were not large enough for a standard deviation focal analysis to detect. The logging roads and patios did have enough variation to be detected. This, although when incorporated with an unsupervised Classification this model was able to detect more detailed land- uses and could be utilized for future small area studies. However, it Should not 17 be utilized for large area studies because it resulted in greater mixing of classes, requiring more editing time. Slope Analysis The second automated model I evaluated was a SLOPE analysis using the DN values of the pixels in the forest canopy. I used SLOPE to identify the rate of change of the DN value from pixel to pixel. Since cryptic deforestation is classed as forest in the 1992 deforestation estimate. I wanted to detect this in the forest canopy, I masked out all land cover in the image except forest. I hypothesized that Since areas of cryptic deforestation are highly disturbed with treefall gaps, logging roads, and patios, the canopy in areas Of cryptic deforestation would have higher Slope percentages than the surrounding undisturbed forest. As per the standard deviation focal analysis, I used TM band 3 in this analysis. I converted band 3 of the image to a grid and in GRID used the SLOPE command to identify the rate of change from each cell to its neighbors, with the result being 3 percentile. The SLOPE function in GRID fits a plane to the values of a 3x3 cell neighborhood around the center cell. The actual algorithm that GRID uses to calculate slope is: rise_run = SQRT (SQR (dz I dx) + SQR (dz I dy)) degree_slope = ATAN(rise_run) * 57.29578 18 The resulting output file was converted to an image file for analysis. To differentiate the Sloped pixel values I color-coded them into intervals of twenty, using the Selection Criteria in the Raster Attribute Editor of ERDAS IMAGINE. When viewing the color-coded result, I could not determine any spatial pattern to map out selective logging areas. I then reclassed the color code, trying intervals of 30, 40, 50, and 60. None of these efforts indicated any Spatial pattern; therefore, this method was also unsuccessful in detecting cryptic deforestation. _l\I_Q_\_I_I The normalized difference vegetation index (NDVI) was also evaluated. Characteristically, green plants strongly absorb visible electromagnetic radiation and strongly scatter near-infrared radiation (Curran, 1980). The NDVI was developed to emphasize the difference between the absorption in the visible and the reflectance in the infrared through mathematical processing of multi-spectral bands, such as ratioing and differencing (Wulder, 1998). The NDVI is a commonly used vegetation index, calculated from the red (R) portion of the visible Spectrum and the near-infrared (NIR) radiance in the form of: NDVI = (NIR - R)I(NIR + R) NDVI has been demonstrated to assist in compensation for changing illumination conditions, surface slopes, and viewing aspects (Avery and Berlin, 1992). 19 After running an NDVI analysis on the image, I ran a texture analysis. This texture analysis is discussed below. I then analyzed the pixel values and removed all undisturbed forest, by the DN value Of the pixels. After removing all undisturbed forest the image was analyzed and I determined that this method was unsuccessful. Like the standard deviation analysis, however, it also detected logging patios and roads. Texture Analysis The last model I tried was a texture analysis on TM bands 2, 3, 4, and 5 of an image to determine if cryptic deforestation could be detected. A texture analysis can be used to segment an image and classify its segments, giving the image Sharper edges. It generally indicates the Spatial variation in neighboring pixel values; further the addition Of texture to an image may add structural information that will aid in the detection Of cryptic deforestation. The first step in this analysis was to mask out all land cover in the image except forest using the 1992 deforestation data. The second step was to, using subset in ERDAS IMAGINE, separate out the TM bands. The analysis was then run on each TM band using a variance algorithm with a 3x3, 5x5 and 7x7 window. The algorithm is: Variance = 2(xii —M)2 n-1 20 where: xij = DN value Of pixel (i,j) n = number of pixels in a window M = Mean of the moving window, where: Mean = 2_"_u n The areas of undisturbed forest were then subtracted from the image. This method was also unsuccessful. Again however, the logging roads and patios were classified as deforestation in TM bands 3, 4, 5 Of the images tested; these were particularly evident in TM band 5. Conclusion Although canopy degradation due to selective logging is visible in the Landsat TM images, it cannot be detected with the any of the above methods. This concurs with the findings of Stone and Lefebvre (1998) who were also unsuccessful at detecting selective logging using automated methods. Although the above methods were unsuccessful at detecting selective logging three of the methods were able to detect logging roads and patios that were previously classified as forest in the 1992 deforestation data. This was especially evident In the texture analysis run on TM band 5. Therefore, I incorporated this method into 21 my research with a buffer routine and a visual interpretation. This method is discussed in more detail in chapter 3. 22 Chapter 3 Methods Employed for Mapping Logging and Degradation Rationale The various attempts to detect cryptic deforestation using automated methods, discussed in Chapter 2 were unsuccessful. Therefore, I incorporated techniques I developed along with methods in previous studies to quantify selective logging in the Legal Amazon basin. Souza and Barreto (1999) developed a method to detect logging patios based on their Spectral characteristics and then used a buffer routine to quantify the area of selective logging in a region in the state of Para. By taking ground measurements in an area of logging, they estimated the radius of logging around a patio to be 180 m. I incorporated the texture analysis on TM band five from Chapter 2 into my study to detect the logging patios and used the 180 m radius to buffer the patios that were detected. This method quantified areas of cryptic logging. Watrin and Rocha (1990) and Stone and Lefebvre (1998) used a visual interpretation to quantify an area Of selective logging. Stone and Lefebvre quantified the area Of logging for a study site in the state of Para. They visually 23 identified areas Of selective logging and manually digitized them on each image in their study. Logged forests were identified by the patterns made by primary and secondary timber access roads and truck loading areas or patios (Stone and Lefebvre, 1998). l incorporated this method into my study, but only digitized areas Of selective logging with obvious canopy degradation. If areas of logging with visible patios and logging roads did not have Obvious canopy degradation, they were not digitized but captured in the texture analysis on TM band five. Data set This study used Landsat 5 Thematic Mapper (TM) satellite images and the1992 Landsat Pathfinder Humid Tropical Inventory data set derived from these images by the project at Michigan State University (MSU), supported by the US. National Aeronautics and Space Administration (NASA). The Basic Science and Remote Sensing Initiative (BSRSI) in the Department of Geography provided the Landsat TM images for this study. As one of the leading institutions Of the NASA Landsat Pathfinder Project, BSRSI has the largest non-govemmental Landsat imagery archive of the tropics consisting of 4000 scenes. The entire satellite data set is referenced by geographic location as well as the Landsat World Reference System path/row footprint (WRSZ). The WRSZ tile system provides an organized spatial structure for data acquisition, cataloging, processing, and overall data management. 24 The deforestation GIS layer was derived from more than 200 Landsat TM images covering the Legal Amazon. The imagery was classified into seven thematic Classes: forest, deforestation, regrowth, cerrado, cloud, cloud shadow, and water. This was done using an unsupervised image classification procedure. Although selective logging is visible in Landsat TM images, no attempt was made to classify the areas of selective logging automatically and it was not included in the 1992 estimate of deforestation. Data Preparation The first step in my research was to examine the more than 200 Landsat TM scenes that encompassed the Legal Amazon Basin at a 1:60,000 scale to identify indicators of selective logging. I identified selective logging areas by first observing a logged area that has been verified in a study done by Uhl (191994) (Figure 3.1). I found thirty scenes that contained selective logging indicators and canopy degradation. The scenes that contained visible selective logging are found mainly in a crescent along the eastern and southern Legal Amazon. These are the scenes that the automated method to detect logging patios identified and on which manual digitizing was done (Figure 3.2) to get a quantification of selective logging in the Legal Amazon Basin. I then had to rectify the thirty images using nearest-neighbor resampling with the four points derived from ephemera data. These points were supplied with the images at the time of ordering and does not take into consideration correct the 25 image to < 500m of the true ground (Chomentowski, personal communication). This was tested at BSRSI by collecting GPS ground points in the Amazon and testing them on several TM images. Automated Detection of Logging Patios In this step of my research I used an automated model to detect logging patios in the forest canopy. The patios were detected using a texture analysis on band five of the TM images. However, so that areas of deforestation and other land- uses did not interfere in the analysis, I masked out all land-uses except forest, using the 1992 deforestation data set. After the patios were detected, a buffer routine was used on them to obtain a measurement of the logged area. The first step was to convert band five of the image to a grid using export in ERDAS IMAGINE. The image may include several bands of information. Each band is a set of radiance values for a specific portion of the electromagnetic spectrum (red, green, blue, near-infrared, Short wave infrared, thermal infrared, etc.) or some other user-defined information created by combining or enhancing the original bands from other sources (ERDAS FIELD GUIDE, 1997). A grid, like a coverage, describes the distribution Of one or more spatial variables. A grid generally describes a Single Characteristic or theme, such as land-use, soils or elevation (ESRI, 1994). Unlike a coverage, which stores geographic information in terms Of lines, points, and polygons, a grid divides space up into discrete units, called cells (ESRI, 1994). A cell has a value describing its characteristics, a size 26 that determines the resolution of the grid, and a position or location defined by a row and column in the grid. The ERDAS image is converted to a GRID format in order to extend all of the GRID GIS software capabilities to the image. This provides access to the actual pixel values in the image. I used band five because in this middle-infrared (mid-IR) TM band, bare soil has a pattern in which high digital number (DN) values are found. This is consistent with relatively high visible reflectance from mineral matter in low organic soils and very high mid-IR DNS in dry soils that have little water to depress mid-IR reflectance (Mausel et al., 1993). In the mid-IR, leaf spectra are dominated by water absorption, giving the forest canopy smaller DNS than bare soil, which is dominated by reflectance. Therefore, the texture analysis is able to distinguish the difference between logging patios and forest using DN values. The 1992 deforestation vector coverage was rasterized using the POLYGRID command in GRID. The POLYGRID command creates a grid from polygons in an ARC/INFO coverage. The 1992 deforestation grid was reclassified into a forest/non-forest map and used to mask the band 5-grid (Figure 3.3). This was done in order to find selective logging Sites and ensure that logging areas already classified as deforestation in the 1992 data set were not Incorporated into this study and accounted for twice. 27 The band five grid was converted back into an image and submitted to a texture analysis in order to detect logging patios. Texture analysis detects spatial variation in neighboring pixels. According to Pratt (1991), many image portions of natural scenes are devoid of sharp edges over large areas. In these areas, the scene can often be Characterized as exhibiting a consistent structure analogous to the texture Of cloth. Image texture measurements can be used to segment the image and then Classify its segments. 1 used a variance algorithm with a 5x5 window. The algorithm is: Variance = 2(xij —M)2 n-1 where: xij = DN value of pixel (i,j) n = number of pixels in a window M = Mean of the moving window, where: Mean = f"_ij Both a 3x3 and a 7x7 window were also tested. The 3x3 window included too much noise, while the 7x7 window excluded too much texture. I then analyzed the texture image to determine the threshold pixel values Of the forest for removal. A portion of the undisturbed forest was selected and the statistics were then calculated for that area. Many randomly selected areas Of undisturbed forest were tested to find the maximum value for removal. The texture image 28 was then converted to a grided data set using the export option in ERDAS. Using a GRID command, the forest was removed from the image, leaving detected logging patios (Figure 3.4). Figure 3.5 is an example of the same image in Figure 3.4 displayed with a color composite of 4, 3, 2, with visible logging patios for comparison. The detected logging patio image was then converted into a polygon coverage. This coverage was then edited to remove all noise but the detected patios. The image with bands displayed at 4, 3, 2, was used to aid in the editing. After identifying logging patios, I used a “buffer” routine to estimate the potential forest area affected by selective logging. In the case of Paragominas, studies within a section Of logged forest indicated an average extraction radius of 180m (Souza and Barreto, 1999). Because this is the only basin-wide detection of selective logging, the 180 m buffer was utilized for the entire basin. Figure 3.6 is an example of buffered logging patios. Digitizing Maps In this step of my study I relied on visual interpretation at a scale of 1:60,000 to identify the location and extent of selective logging from the 1992 images. These areas were verified and digitized manually on each image at a scale Of 1:30,000. The areas were generally found in or around areas of high land-use change and were identified by the characteristic logging patios and logging roads along with 29 Obvious canopy disturbance around the areas. Logged areas leave a characteristic pattern of white points, which are the log landings or patios, embedded in the red hues of the forest canopy. In the areas around the patios canopy degradation may be evident and, if so, was digitized. AS stated earlier in this study, I will refer to the digitized areas as the areas of canopy degradation because these areas not only include selective logging, which is a precursor to fire, but also selective logging areas that have been burned. These burned areas may go beyond the area Of selective logging but are also considered deforestation and result from selective logging. After identifying areas of logging, l digitized these areas into vector GIS layers using ERDAS IMAGINE. The digitizing was done on the very edge of canopy disturbance. l separated the logging areas into two separate vector coverages obvious logging and subtle logging. Obvious selective logging includes spectrally bright patios, roads, and obvious canopy disturbance (Figure 3.7). Subtle logging refers to areas in and around highly logged areas that exhibit obvious canopy disturbance and faded patios and roads, or no patios and roads (Figure 3.8). Logging patios that did not have canopy disturbance were not digitized but were captured in the first step of this research. Because of the 10% overlap of the satellite sensor, some areas of selective logging were in more then one scene, so digitizing was done along with adjacent scenes so that logging areas would not be counted twice. 3O These areas were digitized on the computer screen with the polygons registered to the Universal Transverse Mercator (UTM) raster coordinates of the images. Each x-y coordinate of the polygons digitized was recorded (Table 3.1). 31 Chapter 4 Results Analysis of Total Logging Detected in 1992 Landsat TM Images Thirty Landsat TM scenes with less than 20% cloud cover found to contain selective logging were used to detect selective logging using an automated model following Souza and Barreto (1999). These thirty scenes represent 20% (~1,015,399 km?) of the area of the Legal Amazon and 64% (151,127 km’) of the deforestation. The analysis for the 1992 deforestation estimate Showed that 67% (~3,351,158 km’) of the Legal Amazon was identified as forest and 5% (~237,664 km’) was identified as deforestation due to agriculture. Using the TM band five-texture analysis on the masked images and incorporating Souza and Barreto’s (1999) 180 m radius, I found 1834.001 km2 Of cryptic logging that was not detected as deforestation in the 1992 deforestation estimate (Table 4.1). In order to verify that this estimate did not include areas previously counted in the 1992 deforestation estimate, I took the 1992 deforestation data and the 1992 cryptic logging found in the analysis with ARC/INFO intersect computes the geometric overlay of two coverages, but only those features common to both coverages will be preserved in the output coverage (ARC/INFO, 1994). For the 1992 logging, I found that 26.863 km2 was already included in the 32 1992 deforestation estimate as different classes (deforestation, cerrado, secondary growth, water, cloud, cloud Shadow). l subsequently subtracted 26.863 km2 from the amount Of logging, giving the total cryptic deforestation found of 1834.001 km’. Of the 26.863 kmz, 9.55 km2 is deforestation as agriculture in the 1992 estimate (Table 4.1). Analysis of Canogy Degradation Digitized in 1992 Landsat TM Images The same 30 scenes I used above were interpreted for canopy degradation by means of visual analysis. As discussed in Chapter 3, l digitized canopy degradation into obvious logging and subtle logging. The amount of canopy degradation by obvious logging is 3349.616 krn2 (Figure 4.1) while the amount of canopy degradation due to subtle logging is 1269.359 km2 (Figure 4.2). The total digitized canopy degradation area for the Legal Amazon Basin is 4618.984 km’ (Figure 4.3). Figure 4.4 shows that canopy degradation due to obvious logging is much greater than that due to subtle logging. In order to verify that this estimate did not include areas that had been counted in the 1992 deforestation estimate, the same procedure as discussed above was used. The 1992 deforestation data and the 1992 canopy degradation estimate found in the analysis were intersected in ARC/INFO. I found that 393.724 km2 of Obvious logging was already included in the 1992 deforestation estimate as different classes (deforestation, cerrado, secondary growth, water, cloud, cloud shadow). I thus subtracted 393.724 km2 from the amount of obvious logging to 33 get the total, 3349.616 km’. Of that overlap, 214.464 km’ is deforestation as agriculture in the 1992 estimate (Table 4.1). I found that 82.844 km” of subtle logging is already included in the 1992 deforestation estimate as different classes (deforestation, cerrado, secondary growth, water, Cloud, Cloud Shadow). I thus subtracted 82.844 km2 from the amount of subtle logging to get the total 1269.359 km”. Of that overlap, 43.766 km2 was deforestation as agriculture in the 1992 estimate (Table 4.1). Analysis of Logging and Canogy Degradation in 1992 Landsat TM Images I combined the above two analyses, cryptic logging and total canopy degradation (Obvious logging + subtle logging) to estimate how much selective logging was missed in the 1992 deforestation estimate (Figure 4.5). In ARC/INFO | used the union command, which computes the geometric overlay of two polygon coverages but all polygons from both coverages will be Split at there intersections and preserved in the output coverage (ARC/INFO, 1994). After combining cryptic logging and canopy degradation, and accounting for the overlap in the 1992 deforestation estimate with the intersection command, the total deforestation missed in the 1992 estimate of deforestation was 5308.906 km2 (Table 4.1). Because the texture analysis on TM band five detects logging patios, some logging patios that were digitized into the canopy degradation layers were also detected. This means that part of the total logging estimate of 1834.001 krn2 is also included in the canopy degradation estimate of 4618.964 km’. The union command accounted for this. 34 Using the data from the 1986 and the 1992 deforestation data supplied by BSRSI, I estimated the rate of deforestation to be ~18,000 km2 y“. In Stone and Lefebvre (1998) study, they found that areas of selective logging and surface fires are visible in Landsat TM images but are covered over by regrowing vegetation within one to five years. They found that logging sites in a 1988 Landsat TM image were not visible in the same 1991 Landsat TM image. If I assume that the amount of selective logging found in 1992 was the result of two years of logging activity, then the new rate Of deforestation is ~20,655 km2 y‘1 with selective logging accounting for ~13% of this rate. Comgarison of1991 and 1997 Landsat TM In this study, I compared selective logging on the 1991 and 1997 Landsat TM images for path 226 row 63. I used this scene because Of data availability. I found that in 1991 there was 5.187 km2 of selective logging, and in 1997 there was a total of81.589 km’ of selective logging. In 1997, cryptic logging accounted for 17.514 km2 and total canopy degradation accounted for 74.275 km2 of the overall total. The 1997 Landsat TM image had 20% cloud cover; therefore, logging could be even greater. I found no spatial overlap when comparing the logging areas from 1991 and 1997. Those areas that were selectively logged in 1991 were not selectively logged in 1997. Also, there were no apparent visual clues in the 1997 imagery indicating the logging Sites in 1991. The area 35 selectively logged in 1991 was classified as forest in 1997. Selective logging in 1997 was more widely distributed geographically then it was in 1991. In the Six- year comparison, logging increased 16-fold. Analysis of Basin-wide Verification To verify that all selective logging in the Legal Amazon was captured, I took a random 5% sample of the Images where selective logging was not found. This encompassed 103 images in the western Legal Amazon; five of these images were selected for the verification. Images in eastern Amazon that did not include logging were not used in this verification because the dominant land-cover is cerrado. I ran the automated model on these images to detect logging patios and did a visual analysis to digitize areas of canopy degradation. I did not detect any logging patios in the scenes sampled; in two of the scenes, however, I found a small area of canopy degradation with logging indicators. l digitized these areas and found that 1.08 km2 was missed in path 231 row 67 and 7.581 km2 was missed in path 225 row 65. This equals 8.661 km2 that was missed in my estimate of selective logging. Using this estimate Of logging missed, I calculated that using the methods in this study, I captured 99.8% of visible selective logging for 1992. 36 Analysis of Accuracy Assessment of Logging Patios I conducted an accuracy assessment on the automated model for detecting logging patios by dividing the 30 images in this study into four regions of logging (Figure 4.7). I used the extent Of logging to define the regions and then took a random sample from each region. Two Of the regions were regions of high logging and two were Of low logging. Image path 226 row 63 and image path 002 row 67 were not included in the random sample. I took the random scene from each region (four scenes) and in ARC/INFO built a grid over the image with 18,000 km2 by 18,000 km2 cells. The grid cells were numbered starting with 1, and a random 5% sample was taken from the grid cells for each of the four images. I opened the grid over the images and counted all of the patios within the grid specified in the random sample. I then overlaid the detected logging patio coverage and counted how many patios were detected to get a percent accuracy. The four scenes used in this analysis were path 222 row 63, path 223 row 64, path 226 row 68, and path 230 row 68. In scene path 222 row 63, four cells were analyzed. Only one cell contained logging patios. In this cell I counted one patio and the analysis captured 1 patio. In scene path 223 row 64, four cells were also analyzed. Only one cell contained logging patios, 17 were counted and 16 were captured. In scene 226 row 68, five cells were analyzed. Two of the cells contained logging patios; in the first cell, nine were counted and five were captured; in the second cell, 27 were counted and 24 were captured. In scene 37 230 row 68, four cells were analyzed. Only one cell contained logging patios, two were counted and one was captured. These analyses Show that 84% of the logging patios detected were captured in the TM band five-texture analysis. In this analysis, when counting logging patios, I counted all patios that were visible. Some of the logging patios that were counted were not highly visible and were within areas of canopy degradation where the texture analysis did not detect them. To verify that these areas were included in the estimate Of selective logging, l overlaid the canopy degradation layers onto the Images. So although this analysis Shows that only 84% of the logging patios were detected, some of the patios that were excluded were quantified in the digitizing of canopy degradation. 38 Chapter 5 Discussion The integration Of remote sensing and GIS information has provided important insights regarding selective logging in the Legal Amazon Basin for 1992. I have found that although selective logging is difficult to capture using an automated method, it is possible to quantify selective logging using a visual interpretation combined with an automated detection of logging patios. An important issue was quantified in this study: the extent of selective logging in the Legal Amazon Basin. Results at the basin level Show that ~13% of the rate of deforestation from 1986 to 1992 is due to selective logging. In the 30 Landsat TM scenes where selective logging is found, 19% (~649,499 km’) of the area is forest and 64% (~151,127 km’) of the area Is deforestation due to agriculture. This indicates that selective logging occurs within a proximity of high deforestation. Furthermore, while doing the visual inspection of the Landsat TM images it is evident that the process of selective logging is only occurring in areas of urbanization. The rate logging identified in this study as ~2655 km2 yr1 varies significantly from the Nepstad et al (1999) study where it is estimated that 10,000 to 15,000 km2 39 yr‘ of forest logging is not included in deforestation mapping programs. It also varies, although not as significantly, from a study done by Uhl and Holdsworth where they cite from IMAZON (Instituto do Homen e Meio Ambiente da Amazonia) an estimate that ~8,000 km’ yr‘ is selectively logged. The difference in logging estimates 2,655 km’ yr“1 verses 15,000 km2 W1 respected from my study and the Nepstad et al (1999) study may differ for several reasons. The first is that the Nepstad et al (1999) estimate is based on Interviews of 1,393 wood mills Operating in the Brazilian Amazon and did not use direct observation satellite imagery. The second reason may be because the Nepstad et al estimate is for 1996 and 1997, while my study was done on 1992 Landsat TM imagery and logging rates may have increased. As my analysis on the 1991 and the 1997 Landsat TM image comparison shows, selective logging in one image increased by 16-fold in six years. Nonetheless, this may be an isolated incidence and further investigation in other areas may have a different result. Souza and Barreto (1999) found in their study area that selective logging in 1984 was estimated at 2,089 ha and, in 1996, at 662 ha. The study area Souza and Barreto used has been selectively logged Since the early 19803 this fact may help explain the low level of logging activity detected in 1996. If this is an accurate explanation of the decrease in selective logging for their area, then the Nepstad et al (1999) study estimate of selective logging at 10,000 to 15,000 km2 yr”1 for 1996 and 1997 may be an overestimate of selective logging. 40 Using my data and the data available from the study done by Nepstad et al (1999) I conducted a sensitivity analysis. Using my data I considered the possibility that logging encompasses a larger radius extraction around a patio then 180 m. To test this assumption, l doubled the extraction radius to 360 m, thereby calculating cryptic logging for a larger area. The area captured using a buffer radius of 360 m is 7336 km’. It is not likely that this estimate is higher because the overlap of buffers, and other land Classes, that would decrease this value are not figured into this estimate. Adding this estimate to the area of canopy degradation is 10,811 km’. With the assumption that selective logging is visible for two years the rate of selective logging is ~5,406 km2 yr". Thus the rate Of selective logging per year based on satellite data might range from ~2,655 to ~5,406 km2 yr" (Table 5.1). Table 5.1: Sensitivity analysis calculating a possible lower bound annual logging rates for Nepstad et al (1999) data and upper bound estimate for the data in this study. Negstad et al (1999) This Study Date Reported Based Total 180 m 360 m Estimate on High Canopy Radius Total Radius Total Km2 Intensity Degradation Buffer Buffer 1992 1,737 917 2,655 3,668 5,406 sz yr" Krn2 yr" Krn2 Km’ yr" Km2 Jr" yr" 96/97 10,000- 6,177 1 5,000 sz 41 The data used in the Nepstad et al (1999) study calculated the selective logging rate by determining the log production (m3) in the logging centers. Their estimate was calculated using a range of logging intensities, most of which were high Intensity logging (low-intensity, 19(14-24); moderate intensity, 28(24-32) m3 ha"; high intensity, 40(35-45) m3 ha"). I calculated a high intensity logging estimate on roundwood production with their data to determine the rate if it was calculated with high intensity logging; using their estimate of 45 m3 ha'1 for volume per area. Table 5.2: Calculation of a lower bound estimate of the Nepstad et al (1999) data. Roundwood Low Med. High Logging Assuming Production Logging Logging Logging 96-97 100% (106 m3) Intensity Intensit Intensity (km’yr’l) Intensity % y % % 45 msha yr’1 Total 27.8 49 41 10 9,730- 6,117 km’ 15,090 Thus if we assume that all logging is Closer to the high intensity logging this calculation gives a possible lower rate Of6,117 krn2 y'1 (Table 5.2). Taking Into account my upper bound estimate of 5,406 km2 y’1 there is not much difference in the two estimates. If we also consider the state of Rondonia, where in this study selective logging was not identified whereas, Nepstad et al (1999) estimates between 1,320-1,920 km2 y'1 logged In 1996 and 1997 in Rondonia. Subtraction of the Rondonia estimate from the lower bound estimate of the Nepstad et al (1999) study is 4,497 km’ y", which is lower than the upper bound 42 estimate in this study. Indicating that the Nepstad et al (1999) study did not consider the possibility of counting areas already included in deforestation mapping programs. For instance, in many cases ranchers allow logging on their property before they themselves clear-cut the land for pasture. These instances are included in deforestation estimates. This further indicates that the estimate by Nepstad et al (1999) IS an overestimation. The logging estimate in the study done by Uhl and Holdsworth’s (1997) paper does not state what year or years it’s estimate represents. IMAZON’S estimate is ~8000 km2 yr". The different estimates Show the varying results in this area of concern. It is unknown if any of the above estimates are credible numbers. My study is by no means an overestimate of selective logging, but a Significant indicator that selective logging is evident in Landsat TM imagery and that a large portion of it can be captured as deforestation. More work needs to be done to capture cryptic deforestation. Areas Of new logging may be easy to identify in Landsat TM imagery and captured in this study, but as secondary growth fills in logging roads and patios, the logging Sites become cryptic and harder to identify in satellite imagery. If areas of cryptic deforestation cannot be identified, then there is no way of knowing how much logging was not accounted for as deforestation. Some cryptic deforestation may not be visible in Landsat TM imagery; however, because of the degradation Of the forest, the loss Of biodiversity and the 43 increased susceptibility to fire and carbon release into the atmosphere, the areas that have been missed Should be considered deforestation. Also, small logging Sites are excluded from this study due to the resolution of Landsat TM data (30 m), i.e., they cannot be visually identified. In the study conducted by Stone and Lefebvre (1998), Lefebvre visited the Sites of selective logging, found in their study in Landsat TM images, several years after logging to verify their classification. He found that the intensity of selective logging varied greatly, and the extent of damage to the forest was also highly variable. It was evident that soil compaction by heavy machinery impeded the establishment of new vegetation in roads and patios for several years following harvest. At one site in their study identified as logged, he found that former access roads still had no trees and little other vegetation rooting in the densely packed soil several years after it had been harvested. However, he found that the surface Of the ground was covered with vines and other creeping growth that hid the soil from the satellite. From the ground, Lefebvre identified that dramatic changes in the forest canopy were still evident five years after logging; those trees not logged exhibited some canopy expansion, while fast-growing disturbance-following species (eg. Cecropia spp.), together with vines and understorey growth, combined to form a multilayered and closed canopy. A view of this altered forest canopy from a satellite image would be composed of mixed pixels of partially shaded but vigorously growing vegetation interspersed with the occasional large canopy of a remaining broadleaf tree. The spectral mixture of 44 partially Shaded vigorous growth with the texture provided by these emergent, residual trees, makes distinguishing this from an unlogged forest very difficult using satellite data (Stone and Lefebvre, 1998). Lefebvre found that when viewed on the ground, there was no mistaking a logged forest from an unlogged forest. Automated Detection of Logging Patios Logging patios can be identified in Landsat TM images. The vicinity around these patios is where selective logging occurs, within 180 m in the state of Para according to the study done by Baretto and Souza (1999). The reason I used an automated model to detect logging patios was so I could quantify the area around the patios to obtain an estimate of selective logging. Although Souza and Barreto (1999) determined the Size of the buffer on only 82.5 hectares it was incorporated basin-wide. While visually inspecting the Landsat TM images it was evident that logging patios generally occur In symmetrical patterns and If the Size Of the buffer was increased, as in the sensitivity analysis, a considerable amount of overlap would occur between buffer zones. Also, as shown in the sensitivity analysis even if the buffer was increased in size the selective logging estimate for 1992 would not be as high as current estimates found in studies such as Nepstad et al (1999). In future research it would be beneficial to further test the accuracy Of the 180 m buffer, such as measuring the distance between patios to determine the spatial variability. 45 Some patios in Landsat TM images have evident canopy degradation around them while others do not. I used the automated model to capture areas Of logging that were not evident with canopy degradation to capture the whole of logging impacts 1 digitized in the'areas of degradation. Legal Amazon Basin Digitizing Mags Areas identified as Obvious logging may be areas Of recent logging. These areas have highly visible patios, and the canopy disturbance is identified as lighter Shades of red and darker Shades of gray. Areas identified as subtle logging may be areas Of older logging that burned. These areas had either faded patios or no patios, and the canopy disturbance is identified by Shades of gray in contrast to the surrounding undisturbed forest. Secondary roads and patios are clearly discernible in some of these areas. I was very conservative in my digitizing; consequently, the digitizing of the logged areas may underestimate the selective logging. l digitized the logged areas right on the perimeter of where canopy disturbance was identified. In areas where the canopy disturbance was vast over large distances, I could not determine the edge of logging; therefore, digitizing was done as close to the logging patios and roads as possible. These areas may have been areas of a different terrain or forest type, possibly even areas of large fire disturbances. If primary and secondary roads and logging patios were not evident in the areas of canopy 46 degradation, it was difficult to determine if they were logged areas. If these areas occurred where there was not evident logging, then they were not included in this estimate. This shows that there will be oversights in any large-scale inventory of selective logging in the Legal Amazon Basin using satellite imagery. Considerations about the accuracy of the method During this research project, remote sensing and geographic information systems were combined to provide a quantified estimate of selective logging. I considered three potential problems with the accuracy Of this methodology: logging patio detection, selective logging digitizing, and total area affected by logging. There were two effects that I considered in the detection of logging patios: (1) vegetation recovery and (2) topography. Regrowing vegetation in logging patios quickly covers the bare soil making detection of the patios more difficult with time. In areas of high topography, logging patios may not be detected using a texture analysis because they will be camouflaged by the topography. Visual interpretation is difficult for defining the limits between logged and unlogged forest because disturbance in the forest canopy must be visible, and delineation is considered subjective because it is interpreter dependent. This methodology is conservative and is expected to be an underestimate of the area affected by selective logging. 47 Chapter 6 Conclusion The objective of this research was to conduct the first basin-wide study to quantify the amount of selective logging in the Legal Amazon Basin for 1992 using Landsat TM images. Selective logging is a recent trend, which has researchers concerned. It may have irreversible influences on ecological systems, biodiversity, and carbon release into the atmosphere. This trend, is, in fact, SO recent, that to conduct this research I had to rely on manuscripts in review. I quantified the amount of selective logging in the Legal Amazon Basin using automated model to detect logging patios and by digitizing in selective logging with obvious canopy degradation. This estimate was then included in the deforestation rate from 1986-1992 for a rate of ~20,655 kmzy", with selective logging accounting for ~13% of the deforestation rate in the Legal Amazon Basin. Since selective logging is difficult to detect and regrowth covers the ground in one to five years, the estimate found in this study Should be considered conservative. This estimate is a starting point for future studies that need to be conducted to automate the detection of cryptic deforestation. I have shown that 48 selective logging is difficult to detect using statistical Classification techniques but can be quantified using visual interpretation techniques; visual interpretation, however, is a time-consuming undertaking, especially for large areas. Furthermore, visual interpretation is difficult for defining the limits between logged and unlogged forest since disturbance in forest canopy must be visible and delineation is considered subjective because It is interpreter dependent. Further research in this area is important. Integrating this study with high-resolution satellite information, such as IKONOS with 1 m resolution, would add to the accuracy of estimating logging. Including ground verification using a Global Positioning System (GPS) to reference ground points in satellite imagery would also increase accuracy. Although the methods used in this study are time- consuming it would be beneficial to do this analysis of the 1996 Landsat TM data, to help researchers understand the trend of selective logging, and Its impacts on carbon release, forest biomass, hydrology, biodiversity, and sustainable development. The need to monitor selective logging is immense, and integrating remote sensing and GIS is a useful scientific method for monitoring selective logging and land cover change on a region-wide level. I am hopeful that this research will provide some insight into the detection of selective logging and will help advance future research related to this important issue. 49 Figure 1.1: This is a map of Brazil's Legal Amazon. This study area is an administrative area within the country of Brazil that includes 5x106 km2 of the nine states of Acre, Amapa, Amazonas, Para, Rondonia, Roraima, plus parts of Mato Grosso, Maranhao, and Tocantins. 50 This study area isal cludes 5x10 . Indonia, Roralma. lell’OI The Legal Amazon Guyana Suriname Equador 1' MatO 11"“ GYOSSOI (11‘ LEGEND \ Brazil’s Legal Amazon: Acre, Amapa, Amazonas, Para, Rondonia, plus parts of Mato Grosso, Maranhao, and Tocantis - Brazil outside the Legal Amazon Scale = 1:11,000,000 0 _ I .11.. 200 Miles 5 L Figaro 1.1 51 Figure 1.2: Landsat TM color composite image Of Para, Brazil, for path 222 and row 63, acquired on 24 July 1991. Areas of tropical forest, deforestation, regrowth, logging patios, and canopy degradation. 52 . .,;1 | 1 ' (W ., III‘II‘T I III” Ii'i'iii U I713!” I'jjlmyleIlI ‘ p , I“: 'IIIII IIIIIIIIII ' IIIIIIIHIIII ' IIIIIIIWUI I IIIIIIIIIIII Scale = 1 534,000 2‘" 1"”: ‘. ',( , . (I I'IIN‘ ‘ “J “IIIIII'. “, III“ II 'II’“ Mrfligi In.“ 1.". ‘1' I'lllluh ' "‘ I IIL‘ "“ IIlIIIIIIII J II“ 1.1. IIIIi': :3: 1 I“ ”J“ ‘ It“ [ii/I‘- If“ IIIII‘IIIIIII 7. 'II' , I ‘IIIIIIIIIj‘h , . ”m3 [1 fll ( IIIII Wilt!) I. ”it”. 'IIPI" “I f " ”"II‘II . _ ' III!" ' H“.1, IIIIIJI' MINI]. .IIII-. I III 'UAIEV lIle , .. l M : III, 111131.11)!!!“ v in! 3" m “WI “III.” ‘IIIII. jhlr. i” F a, a F 'R 1‘ III , n E N m .E a, :" IE 5, m o -l G1 .2 H 0 ' 2 C’ W Figure 1.3: This image is a close-up view of figure 1.2, showing an example of cryptic logging and obvious canopy degradation. 54 ' "A?" ’ ’1' ill‘? I ”I, {' ‘IIII . VIIILIIHhIIHh ' “on, . ’ “r”. 1I] III “. III'HIitli‘IiIli ., ”III”. mi. i‘In ‘IIIIIIIIII "I, ;; 'IIII 'I‘Tf‘mii “I III’II'III MI “1' £I|In‘1"£i"kl "II"..“ «I u“. I, .,,,.. . II III.“ ”‘I-mIIIII I II‘II‘IIIIIIII I'i‘I'Iij‘. I'II‘ ‘f'II. I,- \“II 31“” I’d": III" II‘I.‘ I MplIij‘I" ”II, “"I "IIIH'W'III y" . I I“ I ”II “II in; LI , 1);], .‘ . ..‘.I\IIIWI. WIN“ II iIImIIII III IIIgiIIIIIIIII~III.I,. Scale = 1:170,000 F a: a» F 'i a h In a“ h a n. Selective Logg Figure 3.1: Landsat TM color composite image of southern Para state, Brazil, for path 223 and row 65 acquired on 6 June 1993. A selective logging area of Mahogany that Christopher Uhl (199) used as a study site. 56 , “1H“ m. ' ‘ uI‘InIii U :I (I r'" Hui“: _" w ' . .r i. ,' Al“ 1 IS'H’IIIIU- in... . ”t . ~ Imit- "1")" mu 2’“! “'Iiii I ,I-Llf‘fli ~ ”V IL“! , ,. , [III'li _ I . ‘1 I I]; ,1 ,J 1 l' l Christopher Uhl Study Site in Para, Brazil 1993 Scale I 1:118,000 2.6 1.3 0 2.6 Nioneters 57 Figure 3.2: The thirty images that contain visible selective logging and were used in the logging analysis. These images are highlighted in gray and found in a crescent along the eastern and southern Legal Amazon. 58 Amazona a 3? Footprints for Images used in Logging Analysis L.‘ ..... WSII TIIeS/stem shouingpathmdrow nunber Lamps/Longtime Legal A'razon Sate mundaries 400 200 0 200 coordrnatesystem I H Amu fl 1 1: ‘ ‘Tuu‘? Mm ; Gross s:are=1:50.ooo,ooo Snsdcfi Rdedion 400 Klan-tars 0 400 E | ”a i. if? E? iii; = E} E :15! k I [0'0 .39'1'4‘7 10‘ 6i 3% II E. I I E! ‘7' a HM, a‘f ‘3' re" ‘ J J.‘ “_ ages o c 4' or :1 I "E' a 1'}- 5i I l ,. Assam WE]. @ 67 232 68 flit m 3 £335.! n i It m :4 4 “é a a o a" ,0 .4 ‘ I [L'- in n i"- o o o a] o _ I .‘ .I - ‘ HEEL , I; ‘5 3, I I / 5 fr 2 , 'I 1 4 , - o l l I ‘ l I. o ‘0 o o I c h m E? I Figure 3.2 .~ I a. mu]. S9 Figure 3.3: Band five of a Landsat TM image in the state of Mato Grosso, Brazil, for path 227 and row 69, acquired on 5 July 1992. All land cover but forest has been masked out of the image. 60 . iII "IIIIIII nIIIIIl. ‘ ~ ‘ I ._.' I‘ III“. IIIIIIHII Ill! . I “III “III“! 1| "‘II'IIIIIIE “I...“ 1IIIIIIIIII‘IIII4 rIIiII'HuI 'Il'hlI ”I : I VII 4 Irv . H ‘ rIIIIIl}. I1.‘ “l! “ WIIIIII I'll-t-‘II’II “H4 riitlh'” IUI'II .I JHHIIIIHI ” l1: “1.: 61 Masked Image of Band 5 in Mato GroSso, Brazil F's-Ire 3.3 12.8 6.4 N'G‘oro Figure 3.4: An example of logging patios that were detected using the texture analysis on band five of a Landsat TM image. Mato Grosso, Brazil, 19 May 1992. 62 9863:1013 Dug—z «a3 :35 .wmmMGIwfiE :_ .0105. tounEou5< aim: 1.30300 magma— min—mo.— ”.0 63 Figure 3.5: Landsat TM color composite image of Mato Grosso state, Brazil, for path 226 and row 69 acquired on 19 May 1992. This image is the same image as in figure 3.5 showing the visible logging patios that were detected using the automated method. 64 0.5 g . ! sauna: u Slum H Q6 0 '0 0.0 «02. ..Nw..n .ommohmv onus. :. mofifln. 9.500.. ch :o_uN—ufl._mon anocmo 0.....wflmwsnfi 65 Figure 3.6: Landsat TM color composite image of Mato Grosso state, Brazil, for path 226 and row 69 acquired on 19 May 1992. Logging patios that have been detected with the automated method and buffered using the 180m buffer specified by Souza and Barreto (1999). 66 Hun Nmmr ._Nm..m .9320 3a.... :_ :53 5.3 seats: «.33.. 9.39... . o OOOJ-ou r u 0300 . 0.53 n. 67 Figure 3.7: Landsat TM color composite image of Mato Grosso state, Brazil, for path 226 and row 69 acquired on 19 May 1992. Areas of obvious selective logging that have been digitized. 68 iiIlIIIIIII 13. I ”III, I IIHIIII "I? WIIIIII I I IIIIII’ . 7' ‘I IIIIIIII ‘II ,1 _' . I'I‘I. .~ 9 'ii" " «war I IKNIIHII IIWWII" IHIIII“! I. -» IIIIIMIIIv "MI": II. n I” I ’ ”in r- _- L. I -'L‘I ’ If ( MIMI" III" fl”iII‘I'I’IIIII'IIIII’lll, 1 IIHIIIII II III IIIIIIII‘ . ‘ ‘~ I ; " I'll! ”'III“ mm” H li.‘ ‘. 4 . - , H lilu l I . ‘ 2 WI” I . IIIII‘I'I'III II'IIIIIIII'II "IIIII ”III I - -7 , ,IIIII:IIII.-EE r-I lII r' ":2“ I: , l","_..u ”'35.": II» 7“ ‘ " III uh,” 'fl""'IIIII""-' ""III' ‘ * 1,2011%!" : Figura3-7 Obvious Logging in mate Grosso! Brazil 1992 Scale I 1:1 73,900 3.88 1 .93 0 3.88 Kllomotm 69 Figure 3.8: Landsat TM color composite image of Para state, Brazil, for path 222 and row 62 acquired on 24 July 1991. Areas of subtle selective logging that have been digitized. 7O 71 Flgure 3.8 Canopy Degradation Layer B in Para, Brazil 1992 4 Scale 8 1:184,000 Table 3.1: x-y raster coordinates for each polygon that was digitized around areas of selective logging. x-y coordinates for row A is obvious logging polygons. x-y coordinates for row B is subtle logging polygons. 72 Table 3.1 IMAGES 23 063071 1 9223‘ coordlnates t221 t227067 391 470 370 983 371 402 052 359 618 360 819 370 479 632 644 389 959 378 546 71 71 7 162 790 199 701 749 482 906 7 914 131 653 562 304 670 975 681 880 690 101 960 71 897 71 412 741 968 760 673 646 437 659 610 671 476 758 9 14 197 9 978 9 9 494 012 9 490 1 9 9 494 154 9 9 686 9 483 193 9 735 9 775 8 919 463 8 91 779 8 784 744 781 779 779 8 779 1 769 631 8 767 1 8 614 8 084 8 768 937 771 743 767 768 867 8 757 12 841 758 087 8 760 544 745 764 761 478 617 466 096 417 73 9210638 t2260688 387 165 1 67 377 370 194 360 835 401 7 430 740 689 701 469 431 769 778 903 784 839 659 492 664 1 677 1 71 12 717 650 721 727 032 734 1 736 960 7 074 754 659 61 1 10 771 769 438 443 754 447 656 417 750 1 88 637 444 957 000 639 630 1 627 1 627 063 628 002 8 629 001 629 446 627 794 628 1 091 727 183 729 1 739 435 739 041 744 102 7 738 135 751 067 747 727 749 900 1 95 769 11 637 774 14 771 768 984 763 437 644 743 044 709 853 701 134 697 475 683 628 779 1 50 701 058 764 738 059 713 133 1 701 184 701 612 699 153 689 964 690 459 943 722 724 660 768 661 179 673 079 673 524 670 673 524 670 792 1 687 438 687 916 689 701 71 480 721 745 717 055 031 1 400 631 1 119 8 652 86371 8 638 8 629018 677 678 710 8 677 482 678 011 8 677 659 8 67 8 670 895 8 906 681 015 678 892 056 694 434 420 700 994 701 459 699 965 691 694 818 697 153 700 700 727 710 673 701 408 699 436 675 700 703 8 698 840 700 703 698 840 729 713 686 71 746 707 492 704 614 108 714 147 714 173 720 74 077 730 084 694 017 717 750 771 O1 804 533 7 678 750 740 721 820 724 061 71 648 71 714 998 684 948 703 704 659 723 605 727 798 731 094 038 651 629 679 054 730 783 747 630 755 784 780 607 704 730 934 7 914 7 688 801 794 186 149 684 770 711 884 768 086 750 791 12 785 982 703 697 403 811 477 81 1 669 650 649 647 677 910 669 107 677 1 674 437 679 679 1 666 1 61 679 8 601 712 1 452 1 779 681 737 681 1 677 1 1 710 4 71 709 683 707 123 717 009 740 721 751 750 124 8 651 027 638 839 767 1 764 080 760 628 2r t223063A 721 183 706 728 702 393 750 750 759 157 760 030 01 3 771 1 55 794 4 798 692 810 771 753 723 658 465 693 658 620 745 690 71 741 685 070 476 673 994 776 327 688 044 668 944 666 659 794 778 613 744 725 724 405 670 7 41 9 781 4 781 13 823 676 805 770 861 849 838 022 840 665 837 014 864 173 859 438 863 721 14 707 769 8 713 16 8 709 993 717 8 720 567 713 065 710 883 714 928 714 005 717 806 71 817 719 719 644 729 648 401 727 625 728 481 729 421 786 346 740 975 7 160 757 669 756 947 750 175 8 77 930 8 651 758 783 890 714 764 919 070 397 7 953 750 099 747 573 740 678 338 016 600 431 1 673 011 480 929 9 574 15 9 574 040 9 824 9 565 987 9 381 9 399 9 1 994 9 136 75 12230638 7 146 749 817 740 140 720 668 472 393 727 744 924 837 622 827 952 823 951 81 999 875 060 869 183 859 713 833 079 730 894 863 481 870 792 664 110 724 763 330 717 680 151 721 478 764 189 9 9 9 511 979 9 590 429 9 8 400 9 9 584 835 9 569 122 9 9 557 10 9 18 22 Zr t226063A t223004081 0922f 470 817 140 861 764 834 827 635 596 801 318 803 994 81 316 819 908 81 590 826 543 594 856 345 850 818 858 713 836 797 811 722 819 829 419 838 631 839 856 663 625 854 345 843 1 066 837 664 831 403 776 823 893 844 066 821 108 847 991 801 844 756 797 415 189 (0 (0000060010 4 9 507 151 9 503 683 9 14 9 497 723 9 494 9 491 1 9 494 9 9 494 613 9 490 112 9 487 077 9 494 088 9 495 9 4 447 9 4 072 9 4 640 9 4 415 9 471 747 9 467 654 9 459 860 9 466 379 9 439 578 9 424 973 311 180 1 9 407 76 t2230643 185 859 179 865 907 837 839 673 831 835 988 839 723 825 060 827 327 835 905 742 851 19 1 807 946 969 436 679 844 633 045 849 443 849 14 81 756 810 463 81 7 777 185 480 819 099 813 7 800 351 044 855 973 81 823 821 971 848 710 853 039 854 786 849 077 833 665 814 131 ‘D 921 621 738 780 377 308 527 691 791 1 987 (00001000000 1 719 190 1 499 131 9 497 181 9 499 9 9 481 9 4 407 9 4 1 480 9 481 1 477 9 460 9 079 9 450 9 451 874 9 067 9 451 724 9 453 774 807 872 12 419 873 1 1 1 690 977 10 437 740 9440195 20 22 t224063A 1 922r t229070A 6922r t224065A 7 669 090 671 186 158 693 789 689 640 1 15 694 473 854 709 12 71 621 658 396 677 813 707 675 599 985 71 827 164 659 829 300 623 789 1 787 769 858 140 857 19 81 603 567 697 588 930 167 573 017 560 103 1 9 9 9 91 91 387 9177 782 9164448 91 151 91 450 729 41 7 1 84 424 41 421 770 410 409 767 154 8 401 481 8 400 024 7 8 146 9169 3 9170 1 171445 91 935 9137 765 91 333 9147124 9127 1 77 t2230658 t2290708 450 831 485 717 470 651 9 14 169 755 144 547 139 144 t2240653 669 664 668 134 388 660 968 664 4 358 696 1 0 710 458 714 620 71 729 867 750 17 737 71 710 71 380 71 693 693911 670 022 664 752 478 667 479 668 620 670 122 679 420 91 8 699 709 1 401 701 731 709 166 7 977 71 152 711 113 709149 704442 700 71 9 699 767 690 1 74 678 958 434 671 18 630 114 797 774 624 987 621 060 61 190 646 1 747 19 619 15 607 644 604 689 481 607 777 601 006 600 091 604 61 1 788 604 169 608 451 608 969 733 191 78 0260698 71 719 185 717 021 717 171 714 71 4 139 163 664 836 176 648 873 401 642 067 741 127 769 800 175 720 71 1 19 667 674 665 71 714 024 689 1 694 801 717 71 1 621 723 620 13 14 641 11 159 607 442 784 611 10 166 042 770 601 752 195 153 644 9 1 10 663 401 667 420 677 087 759 1 6 694 693 038 691 704 761 180 357 711 482 714 654 71 61 464 044 609 627 059 1 16 61 159 008 661 093 660 742 623 561 581 712 51 193 519 664 540 537 1 540743 619 756 122 728 642 554 098 739 527 524 1 59 659 1 89 553 621 1 642 767 767 645 771 749 476 759 689 7 724 755 025 758 068 7 743 692 71 029 727 161 717 700 400 701 654 690 169 684 721 700 1 704 047 79 t2270688 673 609 51 121 51 523 660 51 7 519 448 624 624 659 405 577 17 4 646 4 345 476 442 924 561 1 12 764 644 006 547 1 O2 708 599 590 433 774 769 767 060 079 760 491 760 174 7 747 102 627 71 1 71 71 711 666 697 441 694 132 1 697 191 8690 8684 654 661 654 708 754 12270708 922r t2250698 21 8922r t228068a 21 t226067051992r t226067a 0 602 499 19 420 421 427 147 409 744 397 014 801 1 197 461 456 349 346 41 170 019 360 137 359 605 409 653 411 366 41 172 417 178 675 41 3 1 52 469 349 461 179 761 730 696 773 657 717 721 190 707 728 710 738 190 701 664 108 1 661 821 909 644 401 644 500 634 182 631 61 1 1 80 t225069b t228068b t226067b 1 772 1 714 161 650 169 1 380 470 017 386 659 1 729 472 1 727 71 714 669 1 7 664 673 1 1 72 1 6 720 620 8 81 092 721 486 8 809 193 8 811 101 12270693 21 8922!“ t2280693 t229069a 21 1922r t2290688 51 627 509 656 927 474 959 470 714 519 521 620 507 429 008 1 185 510 066 51 687 189 7 006 507 499 904 517 140 149 1 146 13 1 459 176 610 1 071 01 3 472 087 743 466 080 991 020 495 099 497 318 8 588 935 500 61 81 t227069b t229069b 530 320 533 661 529 039 544 185 501 495 807 493 412 477 471 479 162 484 085 921 731 499 684 534 527 14 529 097 530 320 523 195 4 329 471 4 1 489 879 480 711 464 547 065 511 757 493 519 488 087 520 494 303 1 470 451 630 931 534 505 193 1 950 61 1 607 17 604 788 842 639 187 607 646 1 1 187 1 772 358 1 12 1 471 172 593 367 14 22 t2290708 6922f t224066a 694 581 12 473 580 1 019 541 050 539 189 531 695 531 537 761 587 579 1 574 708 568 555 638 558 10 558 513 976 516 987 519 379 541 080 539 139 557 739 554 418 550 721 543 034 568 815 570 656 576 1 7 580 121 575 567 074 584 495 466 498 417 747 574 550 085 551 076 565 634 788 559 660 9 085 4 9 093 9 087 4 612 9 081 141 9 134 9 068 795 9 069 9 064 9 063 12 9 054 797 9 058 9 189 490 9 040 419 9 044 371 9 9 9 430 9 039 9 021 479 021 719 9 630 9 023 760 9 019 188 9 031 044 9 028 003 9 627 9 655 9 034 196 9 9 030 744 9 9 031 9 420 9 1 104 8 696 14 166 990 443 985 940 983 419 973 802 452 82 t224066b 461 591 427 13 075 858 799 458 316 547 754 560 141 561 1 9 147 9 597 9 074 498 1 846 9 067 474 9 063 662 9 014 9 952 20 20 5922r t230069a 0922f t2300688 t2230663 734 765 741 165 738 463 760 774 780 575 594 112 601 440 1 17 604 597 14 594 683 600 062 1 21 230 110 231 729 221 225 101 457 247 1 277 434 763 627 154 627 1 12 744 702 086 1 97 471 9 004 643 999 9 001 836 9 83 t230069b t230068b t223066b t222063B 751 109 695 561 564 749 742 574 21 220 658 931 21 789 241 945 244 21 294 660 300 192 294 178 244 384 122 9 169 9 11 653 9 603 729 9 4 151 239 462 733 207112 21 21 849 290 1 411 1 687 1 1 204 042 201622 31 673 649 339 315 274 355 279 089 187 186 898 187 696 186 849 210 608 214 219 395 21 087 437 247 269 271 756 299 054 201 955 1 944 191 807 1 889 486 194 829 968 248 796 31 128 317 963 264 415 261 707 254 466 249 167 1 190 171 190 984 191 209 138 267 188 297 125 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 989 9 498 9 498 529 9 482 676 9 483 631 9 488 871 9 484 934 9 468 707 9 469 679 9 659 9 474 1 9 480 159 9 473 989 9 466 938 9 460 776 84 13 1 969 209 188 220 251 247 647 259 284 603 281 838 330 279 305 244 15 21 1 19 244 250 911 739 1 1 271 748 348 384 194 763 945 254 1 188 418 1 10 438 1 21 220 717 207 194 918 1 719 237 980 605 014 940 685 274 1 60 197 769 310 159 653 § (000000000000 CD 1 644 9 9 9 9 086 9 963 9 176 9 497 9 643 9 494 136 9 481 9 4 01 9 464 9 464 9 9 601 18 947 259 158 1 492 184 091 177 1 181 500 181 110 244 294 328 301 447 219 197 639 21 772 31 301 15 234 744 244 247 241 604 234 7 224 108 224 1 1 19 248 985 254 1 36 101 1 271 003 625 694 239 336 254 263 173 259 718 1 720 716 045 743 743 267 650 9 9 459 9 16 9 470 069 9 470 443 9 1 1 1 9 450 470 9 127 9 447 7 9 1 13 9 9 604 9 1 407 9 760 701 9 7 9 741 476 740 423 9 706 9 709 17 9 704 413 618 9 71 467 9 703 466 9 9 695 9 9 699 750 9 9 690 496 9 687 422 9 753 9 689 166 9 9 422 9 731 9 684 9 427 9 680 13 9 687 9 123 9 444 9 683 440 9 677 768 9 856 9 671 408 9 799 85 t2220628 649 237 1 230 060 725 240 332 239 777 465 786 224 467 736 209 087 209 007 239 045 535 249 678 427 271 462 218 679 21 21 619 127 201 793 227 12 229 231 117 4 344 582 093 9 729 108 697 429 600 694 691 18 9 693 9 687 9 417 9 680 965 9 404 9 1 11 9 650 407 9 857 9 406 9 644 972 9 644 9 639 9 644 9 639 740 9 9 9 1 624 412 9 623 73 9 9 9 1 9 611 481 9 9 271 097 269 257 482 144 7 239 495 237 237 239 19 418 21 015 21 21 21 224 16 1 214 219 684 219 939 214 219 239 680 254 260 257 685 034 5 21 1 21 1 221 163 847 065 847 834 452 809 814 063 81 443 870 870 857 851 9 721 667 9 667 065 9 669 9 670 412 670 747 9 051 9 046 9 9 680 9 679 001 9 679 406 9 668 668 9 668 007 9 677 14 671 143 9 157 9 14 4 9 1 9 651 9 651 9 647 9 661 965 9 659 9 772 9 651 9 649 104 9 7 9 624 9 621 9 630 103 9 9 629 9 630 482 9 621 9 627 363 9 624 633 9 630 068 9 61 097 9 9 754 608 729 9 724 441 9 721 9 9 948 9 711 644 9 968 9 647 9 698 378 9 700 446 9 697 86 t2230628 757 9 144 9 18 747 12 9 7 9 7 727 747 120 7 7 7 057 9 740 173 9 737 9 714 1 7 798 787 779 804 607 937 859 857 128 81 834 417 850 1 859 481 859 871 651 91 858 1 851 616 717 792 837 839 737 837 19 688 849 053 463 967 850 962 857 11 661 851 1 854 720 166 844 1 52 699 724 712 609 731 343 667 731 9 701 082 453 9 690 9 621 9 684 9 448 9 938 9 679 9 327 9 680 437 9 669 9 669 9 455 9 669 107 9 482 9 669 142 671 325 9 659 694 9 651 828 9 647 348 9 647 807 9 1 55 443 9 649 911 9 634 108 9 9 641 9 639 623 9 637 9 634 444 9 857 9 640 666 9 629 9 743 631 9 627 476 9 9 627 050 9 61 644 724 913 9 9 626 857 9 627 9 627 786 9 620 026 9 621 431 9 736 87 159 844 554 829 751 022 751 376 757 103 831 879 874 099 564 716 841 461 17 770 784 809 335 881 647 653 834 161 191 1 877 644 863 092 189 860 115 440 344 854 850 725 854 026 199 414 463 71 884 721 092 619 101 840 603 460 841 192 841 894 632 794 829 122 824 189 863 338 679 728 4 730 639 755 851 422 854 424 9 9 724 10 731 405 71 063 9 717 715 9 691 670 9 699 730 9 731 697 141 864 700 181 9 684 767 9 657 9 9 9 644 9 074 9 634 923 9 9 627 417 9 9 630 1 9 630 9 609 9 609 638 9 61 9 619 9 9 1 9 651 1 649 9 9 8 9 624 700 9 9 61 193 059 9 754 1 9 1 9 744 080 9 741 9 720 9 187 9 719 687 9 71 709 603 9 700 14 700 765 9 701 960 9 700 372 9 701 4 9 704 8972r t22606397A 231 229 880 201 494 4 197 997 248 407 199 742 224 9 t22606397B 9 409 9 61 9 617 9 609 9 779 9 600 054 9 600 9 88 198 090 240 973 21 143 21 462 1 .ANEx wgcsv 058 some .2 @580. 96320 ”:4 059“. 30559". wmaE_ 90%/09.2% Gee/o e6 e0 e0 800 e0e a, «ea/.90? $09.0 ea, 4.02». a, ( ArAv 6A.; Q 61 a; 61 a; 8 Av»? 661 900 «900 909 9090 0(«r90 90( 90( 0909 909 906 90 e904 agape 09.990». 90... 906 90A. 90: 90 9906 9 0 9 0 9 < 9 9 9 4 0 a n. a 0 0 p 9 9 0 p & 0 oow @583 2.230 89 .1 _ 11 m1 .1. 11 1 1 1 _ 1 _ 1 1_ 1o Il1‘IIII 1 1 1 111111. 1 1 1 11 — 1 1 0m 1111 I'l11l1 1 l 1 1i!1l11 I11 11108 09.6 9.6 9.6 9.6 9.6 9.6 9.6 «.96 a6 a6 a6 a6 9.6 96 9.6 9.6 9.6 .296 96 9.6 9.6 9.6 0.96 .ANEx 3:3 6:68 comm .2 9:me 935 ”N6 9599 EOE—=69 mam—E «9 969.69. 00 0000000 9% %49%%%% 9 9 9 9 0 0 0 0 0 ( 99 «999 99 0000000 0000 99%99 99 0« 09 99 «999 99 99 99 om— DON 11 0mm 9:23.. 0:95 awn 90 Total Canopy Degradation IIIII‘II W III I'"I""-"'"I 1,1111 I 'I'1'1'f11111 'I'I'I'1'r" 'II"11I, ""' WWW" "' I'. W 111 "" "II III} I ”'I ”I If I I"? HIKIIIIIé'IaI' "I" III/WI “III "III 1'1 "'I'I'I"1'“ "1'1"" 11'1” III I,.""II"" 'U'IIII""'IIII1""II'II'HIII IIIII III I 1"H' III'I'I'I'III'HIIIII'I1'1'I'I'I'1 ”"1"" ”"11 “MI "£1111,“ “III: 11"“ 11111111111 ' 1 W' “'III% 11 “'IIII‘Il 1: 'IH'IMII :2 "III ”'III n""" "" "I'I"' II ""IIII11""" IIII'I" IIII I 1"HHIIIII" "'II’ 'f'II III "IIII1.'IIIII1‘III1 "IIII 'III "'"HI " 'II '11'I1II‘111I1 '1J'IIII "IIII. "'111'1'1' 1.11.1”11 9'11"”! '1"! "II'I'I' "1'11""I""'." 1111 '."II "’II --‘I.'1"I'I"' 1,1'11'MII I' "III _"'II1 IIIII£I:'”'IIIIII ' "'III.I 1'I1JI I11121I1111' I’m' II II II II 1 I" IIII‘ ""II'E 'IIIII '; "1'1 “1)“"' ”111 II 111,111 11,111..” IIIIIII' """’ "' ""9 'I 1.11. . '1“ 1:: 'IHIHI W' "'H:: “111II III'""’"II ""7! I I. III 'III'I'III"'I1'.‘."1’ III" . 7:1 1”" "'III '1. "-2.111 11 " 11911111 ”I. "I "'IIII'I'III III/III’III I'III." 11'1": III! "I‘I'I'II-I'I'I'Ii I ' III1 II." I"' 1111 "”11"" I 1 M'III Ii: "11"III I- 5.1111 ”1111! ‘.' '1"!- 3 w_ a. ._. —%~ :3 "“1111 1"“ 1 IIIIII1 "'I'1'1'I1 II” 11I111 ”WI" 91 I'IlIlhl ""IIII 1 1"” "IIIII11H'IIIII'I'1'I' “11,11 ’III 1! "I 1"‘113 1IIII1 I "I1 I‘lhh "'""'I )1 II IIIII‘I",,1 III/I. IIII'I'I'1 I‘ 1 "III """1" III III 1' '11‘11 """""" I"""' """III"'I""" 'I'Imw' "'"II ""II ." 1I1'I11 "'II'1" 11 """"1’1'iI'1"""',"I,"3 ' H'III' 111111111111" "‘1' I ""11" "H ”II "I III"? 11 1111”" II‘I'II‘," W" HI I” IIIIIII 1; E11‘flfII '11,. 'III * 1“'I'1 'IIIII "'I"HI1 I" 'II'I~ 111.11 'I "II "'1. "II . 1'1 ""II II. ‘3 “J ‘3 ’\ 9 'b ’\ q « 89 018931919 111.9;919019019 (111,6: 1.90119 18:1 '5 1 15°19- 07${i'35500'j'509’ ““89““ ‘2: 89°89 Image PathIRow Figure 4.3: Total Canopy Degradation (units km’). Obvious logging and subtle logging combined. .rEx 3:3 mcmom zoom .8 9:me 253 new @582 9.6320 #1. 9:9“. U533 255. 9:83 28.30 I _ 33:50.“. 095. 8% 0% «8%, ace, ea, ea, 8% 8% éooéo 0 a; a: o o a. A». 00 ( 09 00 Auqed. (90(00 0 ( a '3“ 0.19 All 00000%0 0 aooa(%9%%aao .1 .11 "111" II" . 1 {P811 111 1 I 11111 11 II 111” l 1’11 1; '1111 111111111 ' '11 1 1 :1 ‘1 1111 I I I 111 I" I 1111 11‘“: 8.: ., I 11 I .11 . I... .l J1 10 . i ’1' 1” II“ 11 111-1 11: 2 1.1 .1 . 1~ I1| I 11' ‘I 11 8 1 .1. 1 [If I"... r 1. . IIIHT I [11 l . . 1|"; 1 cl 8 1 I I 1 11 1 I "1 III ‘11 1 '1'1' 1'1' '1 " "‘11:. 1 1 'I 1 11 I 1111 111 11 '11 1 I x .1 ‘E "'1' "‘1 . 1". 11 1 " 1. I “.111' 9:30.. 23.5 new 9.53.. 2.230 "gown—.9500 Eocuo ea. ea. ea, ea. eel»? will I '.. 8.898889% a0 am..a%9oz A. 0 A. 8 II I ‘1 “11111111111 .1 I 1 com com oov cos 8991 92 .ANEx 3E3 cozfiweoymu B mHmEzmw Nam? 05 E 82:05 go: 9:me 0262mm 9 96 cozmummhohwv Eek ”m6 059m Bom>=mmommE_ 0) a: Av 061 Av Av Au. 62 AV 61 Av Av Av AV 61 Av d. 61 ed. 861 6A»: 58¢ a; 8) Av 61 ea: 861 @661 0e0 000 0000A0 e00»? 00A. 0&0 08 000 00“r 0? 0? o/uc0 0006 00%( 0008% 0800600A 0A8 er 0A fan? 00% 00A. 00% 000 0&6 0Ant 0016 «000 00 0000 0 «0000 0( 00 00 «0 0 0 ( 00 00 0000 0 A 0 a 00 e00 com cow ooov «m2 5 83;. 9.33.. 2 25 5:33.88 .58 oomw 3“” 93 .ANEx 9.5.; 8538 cosmofimmu 30:8 9 92388 02330 cozmumeovou ozabo 6+ 059“. @503 2350! cosmnmamo 30:00. ‘ %o@oo%oo%oe¢o%e o%¢o%¢o%¢o%¢o $¢o¢.¢o,\¢% ea, %c¢o%¢oe¢o %Moo¢&e 00% ea, ooa¢er¢oe¢ooo @00000(00o0:0%00%i00% 0000000 0000 0000000 Bey—Ewan. mum—E 90.0 00; ea. 0‘0 e¢ 00600 .90 0n. 0:9 0 :: Avea/ 006‘ L9 0.: oow oom oov oom coo cos cow :0335500 3050 ucm mEmmo... gumbo com 2"“ 94 Figure 4.7: Images in study divided into four regions by extent of logging for the accuracy assessment of the detection of logging patios. 95 Four Regions Defined for the Accuracy Assessment of the Detection of Logging Patios Legend ‘1 Wriiflinir E x g .. E > . g F N n V g. :3 [ . ‘. § ‘§ § _§ 5 5 8 E " ' § 3 :3 E: '5 E: ( g5 i; 35' g n: I: 2 ¢ ‘1 =5 53d: 3° 5: a a a _&1-"" 'i t" ,2 ° _ .E .E .‘E E E & i557" ":8 z 3: g a «v33 .. 3 .. 3 w g "3'"? 3 W a) g IMII 3‘73 ..... lawn“ i' g M? W" “i W" um “M" Waugh." , I‘ «it 'o 12 ° mm, "““ii 1' ‘M Hm“ W” imp. I W. “"il'i‘rll r. [MN m ‘in W1 “ink: iii-r whim “Mm“ m n "r i‘:‘ i g iq'g‘y’” Win m"‘!ii:‘”‘ I. ”a...“ i L71,” hlil' [11‘ "l ,‘ 1‘12)”: “i lb to 81 {$8 .9 imam, Wm in MI.“ M l’i‘ii’ . I MUM,“ 3“ ’ g g 8 ”1.1;“ in“ W I”. """ili,"‘ 1:. f ‘1‘; 5 8 ‘1‘ ‘ i t v u m V i .‘ K ”:3: is; 3 35m “fig i H 81 ”Flo \ , ‘ .. ‘° 2 ,. 039,335 8‘6] to 5‘5 535 v RI 2k figfig 35$ 8158 $38 a", 5?! “35$ '\ Rafi 81 N g ”5mgfi ”a as Q Ix 35 g g _ {3 g ‘9: ’87 'N‘ V €§Ng§ 8 8 (V g ‘0033 F e N w 8 N MN 04‘ gaggél K .. § § § :tw8§ 888 8_ 353,13 :3 r- Fa‘. ‘— fi 58 0 (ON '- comg <3 ”"3 8N . V .. 8'? " "o‘ ‘2 1, <3 v ‘i‘, h 8'08 DR to 5'08")“, ~ J?“ 8 v k ”803 3333 l?’\% ‘Nm mango f0“ ‘0 3?ng AL} ”Bib 658,3 ,. -°—L_~_~_ (0" w E “,3 a: 00F a 8K8 1!) co 8 z W 0: 30,5 b 8:» z .95 _ ‘8 .3 am TL % 96 Figure 4.7 Table 4.1: Selective logging estimates for each method used in this study. Also. estimates of selective logging that was captured as agriculture in the 1992 deforestation estimate. 2 3 4 6 7 9 Obvious Pathfinder Subtle Logging Total Canopy Cryptic Logging and Logging Captured km2 Pathfinder Degradation Logging Captured Degradation Image km2 in (digitized) Captured in km2 Captured in in 1992 Missed in 1992 1 km2 km2 1 1 .1 . 0.155 38. 10.1 . 1 1 t222062 . . . 1 61 1 . 7 .177 . 1 1 . i 1 .1 . t223062 263.779 12.522 456.973 97.271 0. . . . . 1 14. 1 1 1 . . . 1. . 74 1224067 Column 1 I image Path/Row Column 2 . Digitized obvious logging Includes spectrally bright patios, roads. and obvious canopy disturbance. Column 3 I Digitized obvious logging previously captured as deforestation In the 1986-1992 Pathfinder analysis. Column 4 I Digitized subtle logging includes areas In and around highly logged areas that exhibit obvious canopy disturbance and faded patios and roads, or no patios and roads. Column 5 I Digitized subtle logging previously captured as deforestation in the 1986-1992 Pathfinder analysis. Column 6 I Obvious logging + subtle logging. Column 7 I Cryptic logging captured with the 180 m radius buffer. Column 8 I Cryptic logging previously captured as deforestation in the 1986-1992 Pathfinder analysis. Column 9 - Total cryptic logging and degradation. 97 REFERENCES Avery, T., and Berlin, G., 1992, Fundamentals of remote sensing and airphoto interpretation (5‘h edn.) Toronto: Maxwell Macmillan. Cochrane, M.A., 1998, Forest fires in the Brazilian Amazon. Conservation Biology. 12(5), pp. 948-950. Cochrane, M.A., and Schulze, M.D., 1999, Fire as a recurrent event in tropical forests of the eastern Amazon: Effects on forest structure, biomass, and species composition. Biotropica, 31(1), pp. 2-16. Curran, P., 1980, Multispectral remote sensing of vegetation amount. Progress in Physical Geography, 4, pp. 315-41. Dale, V.H., Robert V. Oneill, Marcos Pedlowski, et al. 1993. Causes and effects of land-use change in central Rondonia, Brazil. Photogrammetn'c Engineering & Remote Sensing, 59(6): 997-1005. ERDAS IMAGINE FIELD GUIDE, 1997, Fourth Edition. Fearnside, p. M., 1992, Greenhouse gas emissions from deforestation in the Brazilian Amazon. In Carbon Emissions and Sequestration in Forests: Case Studies from Seven Developing Countries, edited by W. Makundi and J. Sathaye LBL-32758 UC-402, (Washington: US EPA). GIS by ESRI, 1994, ACR/INFO Version 7. Holdsworth, A. R., and Uhl, C., 1997, Fire in Amazonian selectively logged rain forest and the potential for fire reduction. Ecological Applications, 7(2), pp. 713- 725. Houghton RA, 1995, Determining emissions of carbon from land: A global strategy. In Toward global planning of sustainable use of the earth (Eds. S. Murai). Johns, J.S., Barreto, P., and Uhl, C., 1996, Logging damage in planned and unplanned logging operations in the eastern Amazon. Forest Ecology and Management, 89, pp. 59-77. Lillesand, T., Kiefer, R., 1987. Remote Sensing and Image Interpretation. Toronto: Wiley. 98 Martini, A., Rosa, N., and Uhl, C., 1994, An attempt to predict which Amazonian tree species may be threatened by logging activities. Environmental Conversion, 21. pp. 152-162. Mausel, P., Wu, Y., Li, Y., Moran, E. F., Brondizio, E. S., 1993, Spectral identification of successional stages following deforestation in the Amazon. Geocarto International, 4, pp. 61-71. Mertens, 1997. Modeling of deforestation process in Central Africa: case study southern Cameroon. Applied Geography, 17(2), pp38-57. Nepstad, D. C., Verissimo, A., Alencart, A., Nobres, C., Lima, E., Lefebvre, P., Schlesinger, P., Potter, 0., Moutinho, P., Mendoza, E., Cochrane, M., Brooks, V., 1999, Large-scale impoverishment of Amazonian forest by logging and fire. Nature, p. 505-508. Pfaff, Alexander, 1997, What drives deforestation in the Brazilian Amazon? evidence from satellite and socioeconomic data. World Bank Report. Pratt, W.K., 1991, Digital Image Processing. New York: John Wiley & Sons, Inc. Rankin, J.M. 1985, Forestry in the Brazilian Amazon. In G.T. France and TE. Lovejoy (Eds.). Amazonia, Key environment series, pp. 369-392. Schroeder, P. E., and Winjum, J. K., 1995, Assessing Brazil’s carbon budget: ll. Biotic fluxes and net carbon balance. Forest Ecology and Management, 75, pp. 87-99. Skole, D.L, W.A. Salas and C. Silapathong, “Interannual Variation in the terrestrial carbon cycle: significance of Asian Tropical Forest Conversion to Imbalance in the Global Carbon Budget”. From Asian Change in the Content of Global Change, by J.N. Galloway and J.M. Melillo(editors), Cambridge University Press, 1998. Skole, D. L., and Tucker C. J., 1993, Tropical deforestation and habitat fragmentation in the Amazon: satellite data from 1978-1988. Science, 260, 1905- 1910. Souza, C., and Barreto, P., 1999, An alternative approach for detecting and monitoring selectively logged forest in the Amazon. International Journal of Remote Sensing. (In press). Stone, T. A., and Lefebvre, P., 1998, Using multi-temporal satellite data to evaluate selective logging in Para, Brazil. International Jouma/ of Remote Sensing 19, pp. 2517-2526. 99 Stone T. A., Schlesinger, P., 1992, Using 1 km resolution satellite data to classify the vegetation of South America. IUFRO Remote Sensing & World Forest Monitoring International Workshop, Permanent Plots for world forest Monitoring (IUFRO 84.02.05), Pattaya, Thailand, pp. 85-93. Tucker, C. J., Newcomb, W. W., and Grant, T., 1990, Satellite estimation of tropical deforestation in the Amazon Basin of Brazil. Chapman Conference on Global Biomass Burning: Atmospheric, Climatic, and Biospheric Implications, pp. 15. Uhl, C., and Holdsworth, A. R., 1997, Fire in Amazonian selectively logged rain forest and the potential for fire reduction. Ecological Applications, 7(2), pp. 713- 725. Uhl, C., and Kauffman J. B., 1990, Deforestation effects on fire susceptibility and the potential response of tree species to fire in the rain forest of the eastern Amazon. Ecology 71, pp. 437-449. Uhl, C., and Vieira C. G., 1989, Ecological impacts of selective logging in the Brazilian Amazon: a case study from the Paragominas region in the state of Para. Biotropica 21, pp. 98-106. Uhl, C., Verissimo, A., Mattos, M.M., Brandino, Z., and Vieira, |.C.G., 1991, Social, economic, and ecological consequences of selective logging in an Amazon frontier: the case of Tailandia. Forest Ecology and Management, 46, pp. 243-273. Verissimo, A., Barreto, P., Mattos, M., Tarifa, R., and Uhl C., 1992, Logging impacts and prospects for sustainable forest management in an old Amazonian frontier: the case of Paragominas. Forest Ecology and Management 55, pp. 169- 199. Verissimo, A., Barreto, P., Tarifa, R., and Uhl C., 1995, Extraction of a high-value natural source from Amazon: the case of mahogany. Forest Ecology and Management, 55, pp. 169-199. Watrin, 0.8., and Rocha, A.M.A., 1992, Levantamento da vegetacao natural e do uso da terra no Municipio de Paragominas (PA) utilizando imagens TM/Landsat. Boletim de pesquisa, 124, EMBRAPA\CPATU, Belem, PA, 40. Wulder, M., 1998, Optical remote-sensing techniques for assessment of forest inventory and biophysical parameters. Progress in Physical Geography, 22(4), pp. 449-476. 100 HIGRN STRTE UNIV llllllll Hllll Illlllllllllllllllllllllllllllllll