For each land cover grouping from the 1995 land cover, those pixels were extracted from the 2002 Landsat image. The "expected" class average spectral response was then calculated. The expected value is used to derive a Z-statistic for each pixel. The Z-statistic describes how close an individual pixel's spectral response is to the "expected" spectral response. Pixels that have undergone change between the 1995 land cover image and the 2002 Landsat image will have a high Z-statistic value while pixels that have not changed will have low Z-statistic values. Take for example, a pixel that was classified as forest in 1995 and developed in 2002 would have a large Z-statistic because of the difference in the spectral characteristics between an "expected" forest pixel in 2002 and the existing developed pixel in 2002.
Using visual analysis, extracted pixels were compared to the 2002 image data to determine the threshold between probable change pixels (above threshold) and no-changed pixels (below threshold). Those pixels identified as changed were extracted from the 2002 Landsat image.
This procedure was followed for each of the five land cover goupings.