OSFNet: optical–SAR fusing network for parcel-wise crop classification

Accurate and fine-grained crop classification using remote sensing is crucial for agricultural monitoring and decision-making. Synthetic aperture radar (SAR) data offer the advantage of all-weather, day-and-night imaging, enabling continuous observation of crop growth patterns even under cloudy conditions. However, pixel-level temporal analysis of SAR imagery is hindered by speckle noise, which leads to unstable classification and imprecise boundary delineation. By contrast, high-resolution optical imagery provides rich spatial detail but suffers from limited temporal coverage due to cloud contamination. These complementary strengths highlight the necessity of fusing optical imagery with SAR time series to achieve robust parcel-level classification, ensuring both precise boundary extraction and reliable crop discrimination. To this end, we propose OSFNet, an optical–SAR fusing network for parcel-wise crop classification. OSFNet jointly leverages a single high-resolution optical image and a medium-resolution SAR time series, fully exploiting fine spatial details and temporal dynamics. A multitask learning strategy further optimizes crop classification, boundary detection, and distance regression in a unified framework, and a tailored postprocessing pipeline refines parcel delineation, enabling accurate parcel-level mapping with sharp boundary delineation. Extensive experiments across two study areas demonstrate that OSFNet achieves state-of-the-art performance in multimodal crop mapping, particularly in regions characterized by complex and fragmented field structures. Compared with existing methods, it improves overall accuracy by 1.37%/2.14%, mean IoU by 3.61%/5.10%, and mean F1 score by 1.94%/1.53%.