Abstract:
This raster dataset represents Stand Replacing Disturbance in the Northwest Forest Plan area. Six scene pairs covering the Northwest Forest Plan boundaries were processed to create a landscape-level map of stand-replacing disturbance over the 1998-2003 time period. Decisions made during processing on the amount of change were conservative so as not to omit possible change meeting the "stand-replacing" criterion. Detailed editing using reference photography was not conducted, and spot checking of both dates of photographs for change was not possible, since the 2003 photographs had not yet arrived at the Remote Sensing Lab. Elimination of false change due to moisture/seasonal effects and more detailed change classes requires more classification and editing. This image corresponds to just the first step in our monitoring process.
Supplemental_Information:
Valid Values from the attribute table:
Value ClassName Description
1 tree dv Tree Decreasing Vegetation
5 non-tree Shrub Decreasing Vegetation
10 lnc,unclassified iv, ag, water Area not classified
Imagery
Imagery used was Landsat 5 TM for late summer 1998 and 2003: See nwfp_imagery.xls for image names and dates. The 1998 images were ordered from EROS as terrain corrected, nearest neighbor, UTM zone 10, NAD 27 several years ago for the first North Coast Change Detection Project. The 2003 images are also nearest neighbor, UTM 10, NAD 27 but were only geo-corrected. Images were checked for quality and some bad data was fixed in the 4333 and 4434 '03 images. Terrain correction and coregistration of the time 2 (2003) to the time 1 (1998) images was done in house using the Imagine Landsat model and 30 meter DEMs from the National Elevation Database.
Reflectance calculations
The reflectance calculations were done using Imagine models with text files of image values derived from the image headers and the literature. The 1998 images were done using corrections for Landsat 5 from the Canadian Centre for Remote Sensing ("CCR" model and scripts). The calculations for the 2003 images used the post May 4, 2003 rescaled values for gain and bias for Landsat 5 (to match Landsat 7 as described in http://landsat7.usgs.gov/documents/L5TMCal2003.pdf) with the landsat 7 calculation model ("GAINBIAS"). Our procedures are discussed in the document: rad_reflectance.doc.
Radiometric normalization
Each time one 1998 reflectance image was normalized to the time 2 reflectance image using a linear regression derived from the signatures of (as much as possible) invariant light and dark features.
Scene Processing Areas
A thematic image was developed (sceneproc) using natural boundaries (ecoregions, watersheds) when possible to identify the scene pair used for each pixel. Fog, clouds, smoke (there was no smoke problem in these images), banding, and bad data areas are considered in the scene overlap areas so only the best available image data is used. The sceneproc areas are used as masks when creating the dBGW and PC4 images (see below); the sceneproc image was projected and gridded to create the TM index grid.
PC4 - Computing the change component (4th principal component)
Using an Imagine model, bands 3 and 4 of both image dates.(1998 and 2003) of the normalized reflectance images were stacked into one image. Then the 4th principle component was computed, resulting in a 255 class thematic image. This component correlates highly with vegetation change and closely matches the results of thresholding in most situations. Two images were computed for each scene: tree (conifer, hardwood, and mixed conifer hardwood) and non-tree (shrub, herb, and "barren").
Labeling the PC4 images
Only the classes for larger vegetation decreases (fires and harvests, generally) are being mapped. Usually decreasing vegetation is found in the low-numbered classes, the most extreme change having the lowest class numbers. The center of the histogram correlates with little or no change and the higher numbered classes correlate with increasing vegetation. As a first cut, the PC4 images were "colored up" to visualize the larger vegetation decreases (ldv). To avoid missing any desired change, some classes with less change (ldv+) were kept in this first cut. After mosaicking the tree, non-tree and background (ag, urban, water) images, these ldv+ classes were filtered to keep only the clumps of ldv+ pixels with a minimum mapping unit (mmu) of 11 pixels, approximately 2.5 acres. All other pixels were left unclassified.
This filtered image was edited to remove unwanted pixels (seasonal change, insignificant change, agricultural or urban change for areas that were obviously ag or urban in both dates) or to add missed change. Vegetation change due to fire can be more difficult to label, so .dBGW 8-bit images, clipped to the fire boundary, were thresholded for some of the fires. This thresholding technique is explained in a paper documenting the poster presented at the USDA Forest Service RS2002 conference. The edited image was again filtered to a mmu of 11 pixels.
Mosiacking and final check
The edited, filtered images for the 6 scenes were mosaicked and checked over for quality and consistency. At this stage, the thresholded fire areas were incorporated. Also the scene boundary lines were checked to make sure no change areas split between sceneprocs were eliminated by not meeting the mmu criterion within a scene. Then the mosaicked image was filtered to the mmu.to produce the final image. The final change and TM index grids were projected to teale albers, NAD83.
Analysts Involved
Barbara Maurizi Involved in all phases and prepared the documentation and backups
bmaurizi@fs.fed.us.
Zhangfeng "Leo" Liu Labeling/editing of scenes, terrain correction/coregistration
Pauline Longmire Terrain correction/coregistration