Analysis of four change detection algorithms in bi-temporal space with a case study

Although change detection algorithms for temporal remote sensing images have been compared using various datasets, there is no general agreement on their performance for separating change and no-change. This study compared image differencing, image ratioing, image regression, and principal component analysis (PCA) from a mathematical perspective. Error analysis showed that no-change pixels with errors are expected to be located within an error zone in bitemporal space. Bi-temporal space consists of two temporal axes of target pixel values observed successively. All algorithms confine a no-change area to a zone delineating change and no-change pixels in the space. Image ratioing defines a fan-like sector as a no-change area, generally unsuitable for change detection. The other algorithms confine a no-change area to a strip-like zone. Image differencing defines a no-change zone with a fixed slope, leading to its inability to specify flexibly the error zone that varies with different conditions. In the examined case, image regression and standardized PCA (SPCA) achieved the best performance for change detection, followed by PCA, image differencing, and image ratioing.