ChangeFine: Fine-Grained Change Detection from Remote Sensing Imagery based on Contrastive Prompt Learning with SAM-CLIP

Change detection is crucial for remote sensing applications, widely used in urban planning, environmental monitoring, and disaster response. However, existing change detection methods face several challenges, such as difficulty in distinguishing suspected change features, ambiguous boundaries, and inconsistencies in multi-scale features. To address these issues, we propose a binary change detection method called ChangeFine. This method effectively fuses the detailed features extracted by SAM and the semantic priors provided by CLIP, combined with adaptive mixture-of-experts fusion and change prompt guidance, achieving precise and fine-grained change localization in bi-temporal remote sensing imagery. Specifically, we design the SAM-CLIP Feature Encoder (SCFE), which bridges the modality gap between structural perception and semantic understanding. We introduce the Cross-Perception Feature Module (CPFM), which integrates gated mixture-of-experts and channel–spatial attention mechanisms, enabling fine-grained interaction and adaptive fusion between bi-temporal features, and effectively enhancing the extraction of change-specific features. To handle ambiguous regions (i.e., hard pixels), we introduce the Hard Sample Contrast Module (HSCM), which dynamically identifies highly uncertain samples and helps the model accurately learn discriminative boundaries. Furthermore, inspired by the concept of semantic evolution, we design the Prompt Change Decoder (PCD), which progressively integrates explicit predictions with implicit semantic differences to refine the reconstruction of change regions, achieving a coarse-to-fine change localization. Extensive experiments conducted on four public remote sensing change detection datasets demonstrate that ChangeFine effectively enhances change representation and maintains competitive accuracy and robustness across datasets. The code are available at https://github.com/whudk/ChangeFine.