Patterns in Forest Clearing Along the Appalachian Trail Corridor

Forest clearing in the vicinity of the Appalachian Trail National Park undermines the Trail’s value as a wilderness retreat for millions of annual hikers. We estimate that 75,000 hectares of forest were lost to clearing during the decade of the 1990s inside a 16 km-wide corridor centered on the Trail. This loss represents 2.45 percent of forests within 8 km of the 3,500 km long trail. Managed forest harvests in northern New England accounted for 76.8 percent of forest clearing. The factor most closely related to forest clearing is land ownership: only 0.29 percent of protected forests were cleared, while unprotected and managed forests were cleared at rates of 2.05 percent and 4.03 percent, respectively. A combination of boosted decision tree classifiers, multitemporal Kauth-Thomas transforms and the GeoCover Landsat dataset enabled a single, un-funded analyst to rapidly map land-cover change at 28.5meter resolution within a 3.8 million hectare study area that spanned 16 Landsat scenes. Introduction The Appalachian Trail (AT) is the longest National Park in the United States: a 3,500 km continuous footpath from Mt. Katahdin, Maine to Springer Mt., Georgia. Although famous for its rugged mountain routes, much of the Trail lies nestled among some of the most densely settled counties in the country. A roughly 300 m-wide central path is protected, but in the past few decades the surrounding lands have come under pressure from developers and commercial loggers. The scars left on the landscape from these competing land-uses threaten to undermine the wilderness experience of the Trail’s two million annual visitors, who rely on the AT as a unique forested retreat. The goal of this research Patterns in Forest Clearing Along the Appalachian Trail Corridor David Potere, Curtis Woodcock, Annemarie Schneider, Mutlu Ozdogan, and Alessandro Baccini is to understand the extent and spatial distribution of forest clearing during the decade of the 1990s along a 16 km-wide corridor centered on the AT. Although the official boundaries of the National Park form a corridor that is only 300 m wide, we chose this 16 km-wide corridor in order to better capture the viewsheds of hikers as they traverse the mountain ridgelines of the Trail. For this paper, forest clearing is defined as anthropogenic forest loss, i.e., trees removed for the purpose of timber harvest or land development. Imagery In 2004, NASA released the GeoCover dataset: a global, multidate, co-registered, orthorectified Landsat dataset processed under NASA contract by EarthSat Corporation (Tucker et al., 2004). These circa-1980, 1990, and 2000, Level 1G Landsat images are available free of charge from the University of Maryland Global Land Cover Facility (GLCF; URL: http://glcf.umiacs.umd.edu). Figure 1a illustrates the location of the AT corridor relative to NASA’s Landsat Worldwide Reference System-2 (WRS2). In this discussion, a scene refers to the area of the Earth’s surface indexed by the WRS2. An image is a single data acquisition for a given scene (Strahler et al., 1986). For the AT study, circa-1990 and circa-2000 GeoCover images were acquired at each of sixteen scenes, i.e., 32 images total. GeoCover was essential on two counts: (a) the circa1990 and 2000 imagery is free, and there was no funding source for this project, and (b) the multi-date images of each U.S. scene are co-registered and orthorectified using the same sub-pixel resolution digital elevation model (DEM). Although the GLCF and similar facilities have other circa1990 and 2000 images available for some of the relevant scenes, these images have not been orthorectified using the same DEM and have not been co-registered. Accurate coregistration is essential for change detection (Townshend et al., 1992), particularly in a topographically extreme mountain environment. GeoCover positional accuracy has a root-mean-square error of 50 m and employs horizontal control points for co-registration with sub-pixel accuracy (Tucker et al., 2004). Over the full AT corridor we did not encounter co-registration errors (which would have been apparent in multi-date inspection of the corridor’s mountain ridgelines). To achieve the same geo-registration accuracies as GeoCover over 16 mountainous image pairs would have been very difficult, in part because the image processing PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING J u l y 2007 1 David Potere is with the Office of Population Research, Princeton University, 207 Wallace Hall, Princeton, NJ 08544 and formerly at Boston University, Department of Geography and Environment ([email protected]). Curtis Woodcock and Alessandro Baccini are with the Department of Geography and Environment & Center for Remote Sensing, Boston University, 675 Commonwealth Avenue, Boston, MA 02215. Annemarie Schneider is with the Geography Department, University of California, Santa Barbara, Santa Barbara, CA 93106 and formerly at Boston University, Department of Geography and Environment & Center for Remote Sensing. Mutlu Ozdogan is with the NASA/Goddard Spaceflight Center, Code 614.3, Greenbelt, MD, 20771 and formerly at Boston University, Department of Geography and Environment & Center for Remote Sensing. Photogrammetric Engineering & Remote Sensing Vol. 73, No. 7, July 2007, pp. 000–000. 0099-1112/07/7307–0000/$3.00/0 © 2007 American Society for Photogrammetry and Remote Sensing scheme used to produce GeoCover is proprietary, and the DEMs are not publicly available (Tucker et al., 2004). It is also difficult to mix GeoCover images with any other Landsat product (such as those available at GLCF) because the original images used to create the GeoCover dataset are not freely available. The value of future efforts like GeoCover to provide datasets to be used by many investigators would be enhanced if the original imagery were provided in addition to the final orthorectified products. Although GeoCover has greatly facilitated this research, there are some serious drawbacks that come with constraining an analysis to the GeoCover acquisitions. Figure 1b and 1c characterize the temporal properties of the GeoCover acquisition dates over the full length of the AT. While the images remain free of snow, they are sometimes acquired well after the target periods of peak-vegetation (as late as October). From the standpoint of forest change detection, this causes problems because late images may be beyond the time of year that deciduous trees normally begin to senesce and lose their leaves (leaf-off). In the GeoCover dataset there is also often a poor match between the calendar acquisition dates of the image pairs, with the separation as high as five months (Figure 1b), i.e., causing problems because of shifting solar illumination angles and phenological states between image pairs. The same variability is present at annual timescales; in Figure 1c; 16 AT pairs span intervals from 6 to 15 years. It is worth noting that the variability around the 1990 target date is greater than the 2000 target date (Figure 1c, left versus right). The circa-1990 AT data was collected by Landsat-5, and the circa-2000 data by Landsat-7. The ability of GeoCover to find images closer to the 2000 target year is due to the dramatic increase in the 2 J u l y 20 07 PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING Figure 1. The Appalachian Trail and GeoCover Landsat scenes: (a) Map of the Appalachian Trail and the Landsat WRS2 Grid. The first two numbers of the scene identifier refer to the WRS2 path, and the last three to the WRS2 row. In areas of overlap, scenes were selected which minimized cloud contamination and acquisition date mismatch, and maximized the proximity of acquisition dates to periods of peak vegetation; (b) Image pair calendar offsets for the 16 AT scenes. The 1633 and 1634 image pairs are near-matches, while the 1532 pair is offset by an entire season; and (c) Image acquisition intervals for the 16 AT scenes. The image pair for 1632 is only separated by six years, while 1834’s pair covers 15 years. number of images collected by Landsat-7 as compared with the earlier Landsat instruments, and demonstrates the value of frequent image acquisition and archiving. Fortunately, the wide east-west overlap, or sidelap, between WRS2 paths (Figure 1a) offers a means of mitigating the impact of these temporal inconsistencies in the final map. Because sidelap areas can be covered by either of the two intersecting scenes, the highest quality image pair can be selected in terms of (a) cloud contamination, (b) calendar offset (phenology differences), and (c) temporal coverage. Yang et al. (2001) proposed a similar acquisition scheme for the National Land Cover Database (NLCD) project, using Advanced Very High Resolution Radiometer data to estimate scene phenology. The new MODIS phenology product could likely serve in a similar capacity (Zhang et al., 2003). Remote Sensing Methods Medium resolution remote sensing imagery has been applied to detecting changes in forests for more than twenty years, and a host of viable methods are available (Rogan et al., 2002; Huang et al., 2003; Ustin et al., 2004; Healey et al., 2005). The focus of this paper is not on new methodology, but on innovative application of a suite of existing methodologies over large areas spanned by free datasets, i.e., illustrating that existing methodologies can be used to rapidly produce useful results at low cost. While there is probably no single best approach for change detection, it is crucial to select a remote sensing methodology that is well-suited to both the research question and the available data. To this end, Woodcock and Ozdogan (2004) provide a taxonomy and conceptual model for the change detection process. In their framework, one first builds a description of expected results, and then works backward through the change detection process. Using Woodcock and Ozdogan’s taxonomy, the desired final map of forest clearing along the Appalachian Trail can be characterized as: (a) categorical: the study is focused on a single forest clearing category, (b) spatial: a map is needed because area estimates alone would not answer questions regarding the spatial distribution of forest clearing, (c) local and national: the intended users are both local resource managers (state and county government), and national planners (the Appalachian Trail Conservancy and the National Park Service), (d) decadal: ten years is long enough for detectable change to accrue within the relatively stable forests of the Eastern United States, and not so long that cleared forest stands could recover beyond recognition, (e) endpoint-based: reliance on the free GeoCover dataset restricts this analysis to image endpoints from the beginning and end of the decade, (f) medium spatial resolution: because many patches of forest clearing in the Eastern U.S. are relatively small, it is necessary to work at medium resolution (10 to 100 m pixels), and (g) continental: although the 3.8 million hectare study area would be classed as local in the Woodcock and Ozdogan taxonomy, the Trail’s extreme length makes it a continental feature. These seven attributes of the desired final map inform each stage in the remote sensing process. Multitemporal Kauth-Thomas Transformations The high dimensionality of multi-date imagery poses a challenge for visualizing forest change. Figure 2a and 2b are typical 1990 and 2000 image segments from the AT. Forest change areas are visible as the purple regions in Plate 1c, which is a combination of the original Kauth-Thomas (KT) brightness and the Multitemporal Kauth-Thomas (MKT) transforms of change in wetness and change in brightness (Kauth and Thomas, 1976; Crist, 1985). Collins and WoodPHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING J u l y 2007 3 Figure 2. Spatially aggregated AT forest clearing: (a) Shelter-segment aggregation. Each of these trail segments is defined with an overnight shelter at its center. The coloring of the segments depicts the estimated percent of circa-1990 forest (NLCD) within each segment cleared between 1990 and 2000; and (b) Regularized corridor plot. Each bar is associated with a single trail segment. Bars are ordered as the segments, N to S. The length of a bar (horizontal axis) is the percent of circa-1990 forest cleared within a segment, and the thickness of a bar is proportional to the fraction of circa-1990 forest within that segment. The area under this graph is proportional to the total forest cleared from 1990 to 2000 within 8 km of the AT. cock (1996) first extended the KT transform to multi-date imagery, normalizing a 12-band (two date) transformation matrix to arrive at six new components: change in brightness ( B), change in greenness ( G), change in wetness ( W), stable brightness, stable greenness, and stable wetness. The MKT is well-suited to forest change mapping because: (a) MKT linear transforms are standard from scene to scene, unlike Gramm-Schmidt orthogonalization and principal component analysis; (b) MKT transforms use all of the spectral information in Landsat, unlike most vegetation indices; and (c) MKT brightness, greenness, and wetness bands are physically interpretable, unlike other image differencing approaches (Collins and Woodcock, 1996). In zones of forest clearing, B is positive, G is negative, and W is negative (increase in brightness, decrease in greenness and wetness). Cohen et al. (1998) discuss an application of this approach in the context of mapping forest clearcuts in the Pacific Northwest. It is worth remembering that the Collins and Woodcock (1996), MKT was constructed from the Kauth-Thomas (Tasseled Cap) coefficients for Landsat-5 (Crist, 1985). The GeoCover AT dataset is a mixture of Landsat-5 (circa-1990) and Landsat-7 (circa-2000) imagery. It would be possible to address this issue by constructing a new, mixed Landsat-5 and -7 MKT matrix, using Crist (1985) KT coefficients for Landsat-5 and Huang et al. (2003) coefficients for Landsat-7. We elected instead to use the original Collins and Woodcock (1996) MKT, as the deforestation signal was clearly evident without altering the transform. Viewing the MKT components makes forest loss and gain clearly visible, but it is often difficult to differentiate such forest cover changes from other changes in dynamic land-cover types, such as agriculture. Anchoring the MKT components with a single-date KT component made this differentiation easier. After visual examination across all 16 AT corridor Landsat scenes, a band combination of 1990 B, W, B (RGB) was selected for standard viewing as it proved the most effective at both highlighting forest change and separating that change from other dynamic classes such as agriculture, barren, and urban (Plate 1c). Change Detection Many change detection methods are available that produce categorical results of the kind desired here (forest clearing or non-forest clearing). Decision Tree classifiers (DTs) were selected because they offer a viable supervised method for confronting this study’s large data volume of medium resolution imagery. The 2001 Multi-Resolution Land Characterization Consortium (MRLC) also chose DTs over Artificial Neural Network (ANNs) and Maximum Likelihood classifiers (MLCs) because of DTs ability to handle nonparametric data and their computational efficiency (Homer et al., 2004). DTs have previously been combined with multi-date transforms; Rogan et al. (2002) used DTs with MKT inputs in studying forest change in Southern California, and Levien et al. (1999) have reported success with using DTs and MKTs over larger segments of California. Both Rogan et al. (2002) and Brown de Colstoun et al. (2003) found that DTs outperformed traditional MLCs. Friedl and Brodley (1997) define a decision tree as “a classification procedure that recursively partitions a dataset into smaller subdivisions on the basis of a set of tests defined at each branch (or node) in the tree.” Friedl et al. (2002) have used the C4.5 decision tree algorithm (Quinlan, 1993; source code at URL: http://www.rulequest.com/ Research) in combination with the AdaBoost M.1 boosting procedure (Freund and Scapire, 1997) to yield more accurate land-cover classifications and confidence estimates for the Moderate Resolution Imaging Spectroradiometer (MODIS) Global Land Cover product. Briefly, boosting operates by repeatedly partitioning the training data into training and validation sets, and uses the validation data to focus more attention on difficult cases, thereby “boosting” the overall accuracy of the classifier. Besides incorporating boosting, the BU version of C4.5 is able to read and process digital images. The BU software runs in a UNIX environment, on a SunFire 280r, Dual UltraSparc-III processor with 8 GB of RAM. Individual scene classifiers were trained in an average of 10 minutes, and required between 30 minutes and 1.5 hours to classify the portion of a scene covered by the Trail Corridor. The DTs that MRLC uses for the NLCD are also boosted and MRLC has developed tools to link ERDAS Imagine® with a commercially available, boosted version of C4.5 called C5.0. Separate classifications were conducted for each scene, obviating the need for atmospheric correction (Song et al., 1999). Preliminary work (not shown) suggests that atmospheric correction does not significantly improve classifier performance when the classifier is trained and evaluated on the same scene pair. Prior to classification, a water mask was applied to all 32 images based on Landsat Band 5 (SWIR) thresholds. This water mask included water present in either image in an attempt to minimize confusion between water level changes and forest cover loss or gain; newly exposed shoreline is significantly brighter than water, and hence spectrally similar to forest clearing. The MKT BGW, stable BGW and the single date KT BGW transforms serve as inputs 4 J u l y 20 07 PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING Plate 1. A commercial timber region in North-central Maine (WRS2 1228, 7 9 km area): (a) 1991 image. Landsat bands 4 (red), 5 (green), 3 (blue) color composite; (b) 2001 image. Landsat bands 4 (red), 5 (green), 3 (blue) color composite; (c) Multi-date composite. Color composite of 1991 brightness (red), change in wetness (green), change in brightness (blue). Note that areas of forest clearing appear purple to blue and areas of forest recovery appear yellow; and (d) Final Map. Forest clearing (red), water (blue), and non-forest clearing (white).