Bridging the Intertemporal Domain Shift for Remote Sensing Change Detection via Domain Adaptation

Detecting changes by comparing two remote sensing images of the same area, acquired at different times, is often complicated because these images are typically captured under varying acquisition conditions, including but not limited to different seasons and inherent differences in imaging spectral characteristics. This discrepancy creates a “temporal domain shift,” leading to errors such as false alarms and blurred boundaries in results. To address this challenge, this article proposes an adversarial domain adaptation network for change detection (ADANet-CD). ADANet-CD employs a two-stage progressive optimization strategy. In the encoding stage, it first constructs a dynamic adversarial mechanism between the feature extractor and the domain classifier. At the same time, a temporal mutual-aware attention module (TMAM) is designed to enhance domain-invariant feature discrimination by leveraging the fusion and interaction of spatiotemporal features. Then, a binary-guided distance loss (BDL) based on Minkowski distance is introduced to explicitly constrain the distribution alignment of bi-temporal features. In the decoding stage, an interactive difference enhancement module (IDEM) is proposed. This module aims to sharpen the model’s focus on true changes while reducing its reliance on domain-specific details, further mitigating the domain shift. Experimental results on the CDD, LEVIR-CD, and DSIFN-CD datasets demonstrate that ADANet-CD outperforms existing methods in terms of both precision and $F1$ -score, exhibiting superior performance in both change region localization and nonchange region suppression.