Change Detection Method Based on Pixel Unmixing and EM Algorithm for Low and Medium Resolution Remote Sensing Imagery

The low spatial resolution images often have high temporal repetition rates, but they contain a large number of mixed pixels, which may seriously limit their capability in change detection. Pixel unmixing can produce the proportional fractions of land-covers, on which sub-pixel information can be extracted to reduce the low-spatial-resolution-problem to some extent. Therefore, the sub-pixel change detection can be proposed to overcome this issue in this paper. Firstly, the endmembers of different temporal remote sensing images are extracted; after that, the pixel unmixing process is implemented and the abundance difference image is produced through comparison with the fractional abundances. The endmember variability per pixel is considered during the process to ensure the high accuracy of derived proportional fractional difference image; in the end, an exact threshold should be determined according to the difference image. In order to avoid some false change and noises of the initial fractional difference image, a suitable threshold should be set. In the light of the intensity change values in the image are consistent with the distribution of Gauss mixture, the expectation maximization (EM) algorithm and Bayes discriminant criterion are both used to find out the best threshold of each land cover change. Once the threshold value is set, fraction changes larger than the value will be considered to be correct; other changes will be regarded as zero. The proposed method is compared with two traditional methods in both of the simulated and real experiments. The result demonstrates that our method can extract more precise changed information with high stability.