Improving agricultural land use detection through YOLO–SAM fusion framework

Agricultural land management increasingly depends on accurate object detection and segmentation from aerial imagery. Traditional approaches relying on bounding boxes often lack the precision required for detailed agricultural monitoring, whereas manual segmentation remains time-consuming and costly. This study develops a YOLO-SAM fusion framework integrating the You Only Look Once (YOLO) detector with the Segment Anything Model (SAM) to enhance the accuracy and efficiency of agricultural field detection and delineation from high-resolution drone imagery. The proposed model uses YOLO’s bounding-box centroids as input prompts for SAM, linking semantic recognition with geometric segmentation to generate precise polygon boundaries without manual annotation. The framework also enables direct integration into QGIS for visualization and spatial analysis, thereby supporting practical GIS-based applications. The YOLO-SAM fusion model achieved optimal performance with a confidence score of 0.5 and 80% tile overlap during image tilling, resulting in a detection area ratio of 96.5% and an error area ratio of 4.0%. These results demonstrate the proposed approach effectively manages complex agricultural environments while reducing data preparation costs and providing a scalable, GISoperational solution for automated agricultural field mapping and resource management.