Journal of Korean Society of Agricultural Engineers ISSN:1738-3692(Print) 2093-7709(Online) 한국농공학회논문집

FarmMap is a digital agricultural land map based on high-resolution aerial imagery, containing boundary and attribute information (paddy fields, upland fields, orchards, facilities, etc.) of cultivated farmland, and is utilized in various administrative and research fields such as direct payment inspections and disaster insurance assessments. Since 2022, the transition to a nationwide update system has expanded the scope of work, revealing limitations of the traditional visual interpretation approach, including reduced accuracy, inconsistent quality, and increased time and cost. To address these challenges, this study developed a deep learning-based model using the latest YOLOv12-seg architecture to automatically detect and segment facilities within agricultural land. A dataset of 30,000 image-label pairs was constructed by combining FarmMap and aerial orthoimages, and an effective pre-processing method was proposed using clipping and cropping techniques. Performance comparison across different YOLOv12-seg model sizes (Nano, Small, Medium, Large, Extra-Large) revealed that the Medium model achieved the best performance with precision of 0.921, recall of 0.891, and mAP@50 of 0.946. Notably, by combining deep learning inference results with worker verification in a collaborative workflow, mAP@50-95 improved significantly from 0.812 to 0.961, demonstrating that both accuracy and efficiency in FarmMap update operations can be simultaneously achieved.