Multitemporal Satellite Images for Urban Change Detection

Abstract: The objective of this research is to detect change in urban areas using two satellite images (from 2001 and 2010) covering the city of Shanghai, China. These satellite images were acquired by Landsat-7 and HJ-1B, two satellites with different sensors. Two change detection algorithms were tested: image differencing and post-classification comparison. For image differencing the difference image was classified using unsupervised k-means classification, the classes were then aggregated into change and no change by visual inspection. For post-classification comparison the images were classified using supervised maximum likelihood classification and then the difference image of the two classifications were classified into change and no change also by visual inspection. Image differencing produced result with poor overall accuracy (band 2: 24.07%, band 3: 25.96%, band 4: 46.93%), while post-classification comparison produced result with better overall accuracy (90.96%). Post-classification comparison works well with images from different sensors, but it relies heavily on the accuracy of the classification. The major downside of the methodology of both algorithms was the large amount of visual inspection.