A Multi-Class Crop Field Identification Method Based on Semantic–SAM Fusion and UAV RGB Imagery

Accurate parcel-level crop field information is essential for precision agriculture, field management, and crop monitoring based on Unmanned Aerial Vehicle (UAV) imagery. However, it remains difficult to achieve both reliable crop-type recognition and fine boundary delineation from UAV RGB imagery. Although deep learning-based semantic segmentation models can effectively identify crop types, they often produce coarse or incomplete boundaries. The Segment Anything Model (SAM) can produce high-quality boundaries, but it depends on manual prompts and lacks semantic recognition ability, which limits its use in large-scale automatic mapping. To address this issue, this study proposes a parcel-level crop field identification framework based on Semantic–SAM fusion, enabling automatic semantic recognition and fine boundary extraction without manual prompts. Based on UAV RGB remote sensing imagery, this study developed a two-stage Semantic–SAM framework. Semantic segmentation models, including DeepLabv3+, U-Net, HRNet, and PSPNet, were first used to generate initial results. Then, bounding boxes or internal high-confidence points were extracted from the initial field regions as prompts for SAM to refine the segmentation. The final results preserved crop category information while producing finer boundaries. To evaluate the framework, this study compared four semantic segmentation models and their Semantic–SAM versions on the same-region test set, and further tested their spatial generalization ability on the different-region test set. The results showed that the Semantic–SAM framework provided more consistent gains in boundary quality, with regional recognition accuracy improving in several models and test scenarios. On the same-region test set, the PSPNet-based framework showed clear improvement, with mean Intersection over Union (mIoU) increasing from 78.99% to 83.13% under point-box prompts. The U-Net-based framework achieved the best mIoU of 87.09% with box prompts. On the different-region test set, the DeepLabv3+-based framework showed the largest gain in spatial generalization, with mIoU increasing from 67.22% to 73.45% under point-box prompts. Overall, the PSPNet-based fusion framework showed a better balance in accuracy, boundary quality, and robustness under different-region conditions. These results demonstrate that Semantic–SAM fusion supports automatic multi-class crop field mapping and boundary refinement from UAV RGB imagery without manual prompts or SAM fine-tuning, providing a practical approach for parcel-level crop monitoring and precision agriculture applications.