Using generalized linear models to enhance satellite-based land cover change detection

A popular satellite based land cover change detection technique is used to compare the spectral information for each pixel, from two images acquired at different dates. For each pixel, if there is a big enough difference between the reflectance values from the two images, the area represented by that pixel is considered to have changed. The change detection methods are different in how they determine a "big enough difference". The analyst is left to choose which function of the reflectance values to use and where to set the "change" threshold. These choices are often subjective and affect the accuracy of the change detection. In this dissertation we describe and defend the thesis that Generalized Linear Models can be used to enhance satellite based land cover change detection. Using Landsat Thematic Mapper Data from 1988 and 1994 for an area over Raleigh, North Carolina and a coastal region of North Carolina we evlauate change detection and Generalized Linear Models. For each location, land cover changes are determined from high-resolution air photo reference data. This is coupled with the satellite radiance values for the corresponding area. Generalized Linear Models are then used to regress the binary response of change/no-change (as determined from the air photos) on the radiance values extracted from the satellite imagery. In doing so, the models help determine the most appropriate function of the reflectance values to use for predicting change. For the data in this study, the GLMs indicated a combination of radiance values to be more accurate than a single band or single index. Also, the models indicate that different combinations of radiance values should be used for the different study areas. Next, the models are used to produce "accuracy assessment curves". These curves show the relationship between the location of the "change threshold" and the accuracy of the associated change classification. These curves can be used to compare two models across all possible change thresholds. Finally, the models are incorporated into the satellite imagery to produce "probability of change" (POC) images and "variability" images. In the POC image the pixels contain continuous values ranging from zero to one, representing the probability that the area has changed. The pixels in the variability image contain values corresponding to the variability of the estimated POC.; Results indicate that incorporating Generalized Linear Models into satellite based land cover change detection yields a more quantitative change detection procedure and more informative change detection products. There are three ways to utilize the models. First GLMs can help select the most significant set of explanatory variables to use in the change detection. Next, the output from the GLMs can be used to produce what we will refer to as "accuracy assessment curves". These curves show the relationship between the threshold value used to classify change areas and the accuracy of this classification. The third use is through incorporating the models into the image data to produce continuous "probability of change" images in which the pixel values range from zero to one. These values represent the probability that the area represented by that pixel has changed. (Abstract shortened by UMI.)...