Farmland is a core spatial unit in agricultural production management, and accurately delineating farmland boundaries is of great significance for agricultural resource surveys and cropland monitoring. In recent years, remote sensing imagery has been widely used for farmland extraction. However, traditional semantic segmentation models are limited in their ability to precisely identify semantic contours, making it difficult to effectively extract and interpret the semantic information of cropland from remote sensing images, which in turn affects the statistical accuracy of cropland detection. To enhance the central representation of cultivation semantics, this study proposes a high-precision segmentation method based on the Segment Anything Model (SAM) to address the above challenges. The overall process includes: first, using the YOLOSc network to detect farmland objects and obtain category information; then, using the detection boxes as prompts to guide SAM in performing high-quality segmentation within the specified regions. The YOLOSc neural network is built upon the YOLOv8 architecture and incorporates the C2fScConv and C2f modules to enhance semantic feature encoding and instance-level decoding, respectively. The method was validated on three types of remote sensing data: 1 m resolution UAV imagery, a 2 m resolution mountainous remote sensing dataset, and 3 m resolution Planet satellite imagery. Experimental results show that the proposed method achieves excellent segmentation performance across multiple resolutions, with mIoU scores of 0. 9240 (1 m), 0. 9449 (2 m), and 0. 8211 (3 m), respectively. The method outperforms several mainstream semantic segmentation models on the challenging 2 m mountainous dataset, characterized by complex terrain and blurred plot boundaries. Moreover, the overall method significantly reduces the annotation workload for training data. This offers a new approach for building lightweight and efficient frameworks for farmland extraction from remote sensing imagery. • A detection and segmentation method was designed to introduce SAM into the field of agricultural remote sensing. • In the 1 m resolution remote sensing image, the network can identify more things, and the results are more accurate. • ScConv module is introduced into the network to improve the efficiency and accuracy of the network. • This method used detection labels for training, reducing the workload of data annotation.
Gao et al. (Sun,) studied this question.