Change detection of SAR images based on supervised contractive autoencoders and fuzzy clustering

In this paper, supervised contractive autoencoders (SCAEs) combined with fuzzy c-means (FCM) clustering are developed for change detection of synthetic aperture radar (SAR) images, which aim to take advantage of deep neural networks to capture changed features. Given two original SAR images, Lee filter is used in preprocessing and the difference image (DI) is obtained by log ratio method. Then, FCM is adopted to analyse DI, which yields pseudo labels for guiding the training of SCAEs. Finally, SCAEs are developed to learn changed features from bitemporal images and DI, which can obtain discriminative features and improve detection accuracies. Experiments on three data demonstrate that the proposed method outperforms some related approaches.