Landscape Change in the Highwood Mountains
Methodology:
Source
Data and Classification:
Two image dates, 07/06/91 and 07/05/02, for two Landsat Thematic Mapper (TM) scenes, path 39, row 27 and path 38, row 27, respectively, were used in our attempt to detect landscape change in the Highwood Mountains of north central Montana. The 2002 image was provided through a Forest Service contract by the Remote Sensing Applications Center (RSAC) in Salt Lake. This image was from a different path/row (path 38, row 27) than the 1991 image (path 39, row 27), but due to the location of the study area we were able to work within the overlap between the two images. The 1991 image was acquired through the Global Land Cover Facility (GLCF) at the University of Maryland and was re-projected into UTM, Zone 12, NAD 27 datum, to match the RSAC image. Once this was done, the images were co-registered using the ERDAS Imagine geo-rectification module. After the placement of over 100 ground control points (GCPís), the images were successfully co-registered. Various methods were used to determine the spatial accuracy of each image and in the end it was determined that there was minimal distortion and the pixel mis-match that is often an artifact of the co-registration process had been reduced to less than two pixels.
Once the images were co-registered, separate unsupervised classifications were performed for each image using ERDAS Imagineís ISODATA classification routine. The unsupervised classification routine clusters pixels into classes according to the natural groupings of spectral values that are inherent in the imagery. The user specifies the maximum number of spectral classes allowed and the number of iterations or passes to be made over the image. For the purposes of this study, the intent was to capture the spectral heterogeneity in the image while maintaining the need to keep the total number of classes low. To do this, the unsupervised classification was set in such a way as to produce a total of fifteen spectral classes. Once theses classes were determined, a visual comparison between the unsupervised image and the original raw image was used to decide which spectral classes would be merged with others in order to come up with the final five classes used in the change detection. Using this method, as opposed to simply setting the number of classes in the unsupervised classification to five, allows for more user control in the final determination of classes.
Labeling:
Lack of training data prohibited taking the next logical step, which would have been a supervised classification, and as such the existing classes determined by the unsupervised classification were labeled by visual comparison with a Digital Ortho-photo Quad provided by the Forest Service. The labeling process proved to be hindered by some confusion between the unsupervised classes, so in an effort to achieve better pattern recognition and less spectral confusion we removed all spectral classes associated with conifer (the dominant type and thus the source of the confusion). This was performed by exporting the unsupervised classification to a grid, editing out all conifer classes in ARC/INFO, and then using the resulting file to mask out all of the pixels associated with conifer in the raw image. Once this was done, another round of unsupervised classifications were performed on the raw imagery (sans conifer). It was hoped that by removing such a dominant spectral class the ISODATA routine would have greater success at segmenting the image in accordance with the remaining pixels in the raw imagery. We believe that this operation was successful in that the resulting classes appear to more closely resemble the patterns that exist in the imagery and on the ortho-photo quad. However, again, labeling success remained limited by the paucity of point data.
Results:
