Unsupervised Spatial Feature and Change Detection in RS Imaging

We adapted and completed the spectral unsupervised clustering algorithm in terms of modern high-dimensional nonparametric density estimation methodology. This led to the completion of the unsupervised spectral classification part of our system. We then studied possibilities to improve our method of geo-spatially biased sampling of pixels. One of these techniques, based on a Bayesian geo-spatial local/global density quotient seems to be the most promising to provide efficient spectral samples for the ensuing, second, unsupervised spectral classification step. Finally, we completed the third step, the allocation of all pixels in the image to the system of classes in the second step in terms of two optional methods.