Semi-Supervised SAR Image Change Detection with Complex-Valued Graph Contrastive Learning

Multitemporal change detection in synthetic aperture radar (SAR) images poses a significant challenge due to severe speckle noise and limited availability of high-quality labeled samples. To address this issue, we introduce an effective semi-supervised change detection network known as Complex-valued Graph Contrastive Learning (CGCL-Net). This approach fully leverages the amplitude and phase information in SAR data, extending the change detection model to the complex-valued domain and capitalizing on SAR image advantages. Specifically, we employ an unsupervised complex-valued graph representation learning module to extract meaningful graph-level representations from complex SAR data. Additionally, a semi-supervised complex-valued graph-based change detection model is proposed for predicting change detection results with a small number of embedded labeled samples. We evaluate the performance of our method on two multitemporal SAR image datasets and compare it to other image change detection algorithms. Experimental results clearly indicate that our approach surpasses other state-of-the-art methods in SAR image change detection.