作者
Jingxiang Yuan,Xiaobo Tan,Gufeng Gong,Guoxiong Zhou,Ming’e Zhu
摘要
High-resolution forest change detection (FCD) is essential for carbon assessment and sustainable forest management, yet reliable detection in complex multi-temporal scenes remains difficult because phenology-induced pseudo-changes, irregular and blurred boundaries, and unstable model adaptation under limited supervision jointly degrade performance. In this work, we propose Temporal and Boundary Refined Mamba (TBR-Mamba), a Siamese state-space framework tailored for high-resolution forest change detection. To support this task, we further construct HNF-CD, a public, large-scale, high-resolution benchmark for complex multi-temporal forest scenes, featuring expert-guided annotation of semantically verified cross-category transitions while retaining phenology-related and other non-semantic hard negatives. To handle temporal discrepancies, the Phenology-Resistant Temporal Contrast (PRTC) module suppresses phenology-induced pseudo-changes through temporal contrast-guided attention and asymmetric refinement in a dual-stream interaction scheme. To resolve spatial ambiguity, the Edge-Prior Guided Decoupling (EPGD) module sharpens blurred boundaries by decoupling boundary cues from region semantics under boundary-aware joint optimization. Finally, to stabilize convergence under limited annotations, the Hierarchical Transfer Fine-Tuning Encoder (HTFT-E) adapts the Swin-UMamba backbone through stage-wise transfer control, promoting effective transfer of generic visual priors to forestry-specific semantics. Experiments on HNF-CD, Forest-CD, and Zhejiang-CD show that TBR-Mamba achieves the best overall performance among the compared methods, reaching F1 / cIoU scores of 89.81% / 81.50%, 95.14% / 90.74%, and 92.60% / 86.23%, respectively. Cross-dataset transfer experiments further suggest that HNF-CD serves as a useful forestry-specific intermediate source, complementing generic ImageNet pre-training. Overall, these results indicate that TBR-Mamba is effective for fine-grained forest change detection in heterogeneous forest environments.