Change detection in unlabeled optical remote sensing data using siamese CNN

In this article, we propose a new semisupervised method to detect the changes occurring in a geographical area after a major damage. We detect the changes by processing a pair of optical remote sensing images. The proposed method adopts a patch-based approach, whereby we use a Siamese convolutional neural network (S-CNN), trained with augmented data, to compare successive pairs of patches obtained from the input images. The main contribution of this work lies in developing an S-CNN training phase without resorting to class labels that are actually not available from the input images. We train the S-CNN using genuine and impostor patch-pairs defined in a semisupervised way from the input images. We tested the proposed change detection model on four real datasets and compared its performance to those of two existing models. The obtained results were very promising.