Time-Series SAR Image Change Detection via Graph Transformer with Contrastive Learning

Abstract:
Time-series SAR images exhibit spatial and temporal correlations, providing richer information about temporal variations. However, effectively utilizing the information from time-series images is a significant task in the current change detection field. To address the above challenges, this paper proposes TCNet, a novel joint Time-series SAR image Change detection network that combines graph Transformer and Contrastive learning. Specifically, we introduce a time-series graph construction module that enables the perception of spatial-temporal correlations within the time-series SAR data. This module facilitates the understanding of how spatial and temporal factors are interconnected. Additionally, we propose a new time-series graph representation learning model that creatively combines graph Transformer and contrastive learning, named SGCLformer. SGCLformer aims to enhance feature representation’s differentiation ability and robustness, enabling extracting the potential change information in the time-series images. Finally, to evaluate the performance of TCNet, we conduct experiments on two time-series SAR image datasets. The experimental findings highlight the effectiveness of TCNet in achieving superior change detection performance.