Few-Shot change detection in optical and SAR remote sensing images for disaster response

• A dual-stage training framework addresses few-shot change detection between optical and SAR images. • A novel module enables cross-task feature transfer for enriched target task representations. • Extensive experiments demonstrate superior performance, strong robustness, and effectiveness in complex scenes. • Strong generalization is validated in real-world earthquake and flood scenarios. Few-shot change detection in optical and Synthetic Aperture Radar images is a critical task for disaster monitoring. offering significant application value in complex scenarios with extremely limited labeled samples. However, the randomness of disasters causes a notable data distribution shift between public datasets and real disaster scenarios. With only a few annotated image pairs, existing methods struggle to effectively fuse features from heterogeneous images, leading to severe performance degradation. To address this challenge, we propose a Dual-Stage Training framework for Change Detection (DSTCD), specifically designed for few-shot scenarios involving fewer than 20 labeled image pairs. DSTCD first undergoes source task pre-training on a heterogeneous image registration dataset. Subsequently, in the target task stage, it leverages task guided feature transfer module to transfer the structural and semantic features of image registration to the change detection task. This mechanism significantly enriches the feature representations under few-shot conditions, enabling accurate identification of affected areas. To validate its performance, we conducted comparative and ablation studies against eleven state-of-the-art methods on four public datasets covering both urban expansion and water expansion scenarios. Experimental results demonstrate that DSTCD achieves a significant performance lead. Its average F1-score surpasses the second-best method by 6.98% in urban expansion scenarios and by 13.09% in water expansion scenarios, proving its superior performance and strong multi-scenario adaptability. Furthermore, robustness analysis of varying training sample sizes and real-world disaster application validation further confirm the method’s practicality and robustness for data-scarce disaster monitoring tasks. The code of the proposed method will be made available at https://github.com/Lucky-DW/DSTCD .