Adversarial Mask-Guided Generation for Multi-Temporal Change Detection in Remote Sensing

Accurate change detection in multi-temporal remote sensing imagery is critical for applications such as urban development monitoring, disaster response, and environmental surveillance. However, the performance of deep learning-based models is often hindered by the scarcity of annotated image pairs. Although various semi-supervised and self-supervised methods have been proposed to alleviate annotation requirements, they often lack fine-grained control over the spatial and semantic consistency of changes. To address this limitation, we present GenCD-GAN, a novel generative adversarial framework tailored for synthesizing realistic and semantically consistent change pairs. Unlike conventional GAN-based methods, GenCD-GAN introduces an adversarial consistency loss to preserve structural integrity, an adaptive change masking mechanism to refine fine-grained changes, and a multi-modal augmentation strategy to enhance robustness under diverse conditions. Extensive experiments on the LEVIR-CD and WHU-CD datasets show that GenCD-GAN consistently improves the performance of state-of-the-art change detection models by generating high-fidelity synthetic training data. Compared to existing approaches, GenCD-GAN offers better interpretability, stronger semantic alignment, and controllable change synthesis. The proposed framework offers a scalable and effective solution to reduce manual annotation efforts and advance deep change detection systems in remote sensing.