SAM2-CD: Remote Sensing Image Change Detection With SAM2

Change detection in high-resolution remote sensing imagery remains challenging due to the difficulty in distinguishing task-relevant semantic changes from irrelevant variations and capturing subtle local differences. While Segment Anything Model 2 (SAM2) exhibits strong generalization in natural image segmentation, its direct application to remote sensing change detection is hindered by single-image segmentation bias and contextual granularity mismatches. To address these limitations, we propose SAM2-CD, a lightweight architecture that adapts SAM2 for bitemporal change detection through two novel modules: An Activation Selection Gate (ASG) that dynamically suppresses task-irrelevant variations by learning channel-wise activation maps from cross-temporal features, and A Global-Local Contextual Attention Module (GLCAM) that hierarchically integrates adaptive pooling and spatial attention to amplify both scene-level semantics and pixel-level details. By leveraging SAM2's multi-scale pyramid encoder and our optimized Multi-Scale Feature Fusion Module (MSFFM), SAM2-CD achieves state-of-the-art performance across three benchmarks (LEVIR-CD, WHU-CD, LEVIR+-CD), with IoU scores of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$85.51\%$</tex-math></inline-formula>, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$88.97\%$</tex-math></inline-formula>, and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$69.31\%$</tex-math></inline-formula>, respectively. Notably, cross-dataset experiments demonstrate superior generalization, outperforming baselines by <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$35.29\%$</tex-math></inline-formula> in F1 under zero-shot settings, demonstrating superior accuracy and robustness in complex scenarios.