Enhancing Segment Anything Model for remote sensing change detection via spatial–temporal adaptive fine-tuning

The Segment Anything Model (SAM) has emerged as a groundbreaking foundation model for general-purpose image segmentation, demonstrating exceptional zero-shot generalization capabilities in conventional vision tasks. However, its direct application to remote sensing analysis reveals significant performance limitations, particularly in change detection. This task presents critical challenges as it requires both identifying changed regions and categorizing semantic transition types. The reason SAM cannot be directly applied to remote sensing change detection is that it fails to effectively extract temporal change information, which is essential for accurately identifying and categorizing changes over time in remote sensing imagery. To address this domain adaptation challenge while maintaining parameter efficiency, we propose a dual-module fine-tuning framework that preserves SAM’s pretrained knowledge by keeping the encoder parameters frozen while introducing two specialized adapters. First, the SpatialLoRA module enhances multiscale feature extraction through hierarchical low-rank adaptations, which are specifically optimized for high-resolution remote sensing imagery. Second, the temporal-change adapter module captures spatiotemporal dependencies in bi-temporal image pairs via learnable temporal fusion operators. In addition, we propose the TranScd decoder, a multistage decoding framework comprising three cooperative modules that systematically capture hierarchical spatiotemporal dependencies to jointly handle semantic change detection (SCD) and binary change detection (BCD). Comprehensive evaluations on benchmark datasets (WHU-CD for BCD, SECOND for SCD) demonstrate state-of-the-art performance, surpassing existing methods for both SCD and BCD tasks. Notably, our approach attains these advancements with only 10.3 M trainable parameters, surpassing conventional parameter-efficient fine-tuning techniques by 7.63% to 11.55% (F1-score) and 5.92% to 8.21% separation kappa coefficient (SeK) in accuracy under comparable parameter budgets. These results validate that targeted architectural adaptations effectively bridge the domain gap between natural and aerial imagery while preserving the foundational strengths of vision foundation models.