RFHP-CD: A Prompt-Driven Fine-Tuning Framework of Remote Sensing Foundation Model for Building and Cropland Change Detection

With the widespread adoption of foundation models in computer vision, remote sensing foundation models have emerged as a key approach for advancing intelligent Earth observation. However, most existing foundation models are pre-trained on natural images, leading to semantic gaps and task mismatches when transferred to remote sensing downstream tasks—especially change detection, which requires fine-grained spatiotemporal modeling. To address these challenges, we propose Remote sensing Foundation models’ Hierarchical Prompt fine-tuning framework for Change Detection (RFHP-CD). Built upon the state-of-the-art remote sensing foundation model HyperSIGMA, RFHP-CD introduces a lightweight and efficient Tri-prompt module, which employs a tri-level prompt encoder and a hierarchical prompt interaction strategy. In addition, we design a Multi-Stage feature Difference modeling and Fusion (MSDF) mechanism to guide the model in deeply capturing bitemporal change information. Experiments conducted on three representative datasets—LEVIR-CD,WHU-CD (building/bare land change) and PX-CLCD (cropland change)—demonstrate that our method outperforms mainstream approaches across multiple evaluation metrics, achieving F1 scores of 91.15%, 91.39% and 73.88%, respectively. Ablation studies and t-SNE visualizations further verify the effectiveness of our modules in enhancing model discriminability and preserving feature separability. This work offers new insights and practical pathways for adapting remote sensing foundation models to downstream change detection tasks.