A Comparison of Four Algorithms for Change Detection in an Urban Environment

Abstract Four digital change detection algorithms are applied to 1986 and 1990 Landsat Thematic Mapper (TM) images of a portion of the Salt Lake Valley area to determine the land-cover/land-use changes between the two dates. Image differencing and image regression are used with the six reflective TM bands to create 12 change images. A tassled cap transformation is also used to create three change images (change in brightness, greenness, and wetness). A new method—a Chi square transformation—is proposed and used with the six reflective bands to create a single band change image. A thresholding strategy is applied to the change images to separate the pixels of change from those of no change. Five hundred eighty-five samples are selected through a combination of stratified random sampling and systematic sampling procedure. Ground truth information on the sample sites is obtained from the interpretation of color aerial photo slides of the two dates. Three indices are used to assess the accuracies of the sixteen change images for land-cover/land-use change detection. The regression of TM Band 3 is found to be most accurate for detecting change vs. no change in all three indices, while the difference image of TM4 is found to be least accurate. The kind of change in land-cover/land-use is also examined. The results are compared and summarized. Changes involving construction sites and farmlands are found to be accurately detected by several change images.