Neighborhood Difference-Based Self-Supervised Network for Detecting Small Changes From Synthetic Aperture Radar Images

Change detection from synthetic aperture radar (SAR) imagery is a significant yet challenging task due to the intrinsic speckle noise and different object appearances in SAR images. Most deep learning-based techniques rely on sufficient labeled samples as supervision for training and lack the ability to identify subtle changes. In this letter, we propose an unsupervised neighborhood difference-based self-supervised network (NDSNet) to detect small changed areas in SAR images. Specially, an antinoise neighborhood differencing operator is designed to generate an initial difference image for segmentation into changed, uncertain, and unchanged classes. Then, a simple pseudo-Siamese network consisting of an encoder and a predictor is established to extract discriminative features from raw data based on contrastive learning, which takes unchanged paired image patches as input and is trained in a self-supervised learning (SSL) manner. On this basis, the encoder followed by two fully connected layers constitutes a classification network, which is trained by the samples belonging to changed and unchanged classes. Finally, the uncertain samples are assigned changed or unchanged labels by feeding them into the network, thus obtaining a binary change map. Experimental results on three SAR datasets confirm the superiority and robustness of the proposed method, achieving an overall accuracy of up to 99.8% in detecting small changed regions. Our codes are available at
https://github.com/OMEGA-RS/NDSNet
.