BiSAM-CD: Zero-Shot Remote Sensing Change Detection via Bidirectional Temporal Memory in SAM2

The reliance on annotated training data remains a critical bottleneck for remote sensing change detection, particularly in scenarios requiring rapid response. To address this challenge, our framework operates in a fully training-free paradigm that leverages the zero-shot (ZS) generalization capability and temporal memory mechanism of the segment anything model 2 (SAM2). By redefining change detection as a bidirectional object tracking task, our method constructs ascending and descending frame sequences with adaptive interpolation to mitigate environmental noise. Two lightweight modules are introduced: Module A suppresses pseudo-changes through cross-sequence mask validation, while Module B stabilizes tracking via nonlinear frame interpolation. Evaluated on WHU-CD, LEVIR-CD, and SECOND datasets, our approach achieves $F1$ scores of 85.31% (mask prompts), 57.43% (SAMPoly prompts), and 39.35% (auto-masks), respectively, rivaling supervised models like FC-Siam-Cone (89.53%) in building change detection while demonstrating ZS generalization to multiclass environments (SECOND dataset) without requiring semantic transition modeling. Ablation studies confirm the necessity of each module [temporal object tracking consistency and cross-sequence change verification (TOTC-CSCV) and bidirectional frame sequence construction with adaptive interpolation (BFSCAI)], with combined contributions improving $F1$ by 27.17% over the baseline. The framework processes $1024\times 1024$ images in 0.0325 s (Base+) with <1 GB memory, achieving $4\times $ speedup over supervised counterparts. The framework demonstrates practical potential in resource-constrained environments, showing particular promise in rapid disaster assessment (WHU-CD) and multiclass land cover monitoring (SECOND). Its dependency on external prompts is mitigated by compatibility with diverse sources (OWLv2, SAMPoly, and SAM2 auto-masks), though autonomous prompt generation remains future work. The code will be open at: https://github.com/KimotaQY/BiSAM-CD