Ensuring national food security and sustainable agricultural development requires accurate and timely detection of cropland changes from remote sensing imagery. The detection of land use changes is more challenging than traditional change detection due to the presence of scattered land parcels and similar spectral information. In this paper, we propose CroplandSAMCD, a Siamese change detection framework that integrates the visual foundation model MobileSAM with a fine-tuned TinyViT encoder and a domain adaptation fusion module (DAFM). MobileSAM extracts high-level semantic features, while TinyViT refines fine-grained boundaries of small cropland parcels. DAFM hierarchically fuses semantic and detailed features, mitigating domain discrepancies between natural and remote sensing images. The experimental results on the Cropland Change Detection dataset (CL-CD) cropland data set show that the proposed CroplandSAMCD model achieves a good balance between accuracy and efficiency. Specifically, this method achieves an F1 score of 81.99%, which was 4.27% higher than the previously best method. Further comparative results demonstrate that CroplandSAMCD outperforms several benchmark methods proposed in recent years in terms of overall accuracy and boundary integrity, verifying its effectiveness and competitiveness in the task of cropland change detection. In the generalization experiments, CroplandSAMCD demonstrated excellent performance on the WHU-CD and LEVIR-CD building change detection data sets, fully showcasing the robustness and universality of CroplandSAMCD.
Li et al. (Wed,) studied this question.