Paddy parcel detection method using drone images based on YOLO-Seg model and tiling movement

Tropical Plant Biology · Springer Nature · 2026-09-02 · DOI: 10.1007/s10333-026-01092-5

Аннотация

Given the increasing need for accurate agricultural land monitoring due to the discrepancy between official records and actual land use in South Korea, an urgent demand is for automated monitoring systems. This study aims to develop a system that combines object detection and segmentation capabilities in high-resolution images using the YOLO-Seg model for paddy parcel monitoring based on drone imagery. We propose a centroid-based tiling-movement technique together with a systematic analysis of optimal overlap rates (0–90%) to reduce tile boundary errors that occur in large-scale drone image processing. At the tile-level training stage, the YOLO-Seg model achieved 84.5% precision, 62.8% recall, and a mean Average Precision (mAP) of 62.1% at an IOU threshold of 0.5; the comparatively low recall reflects the fact that paddy parcels truncated at tile edges are penalized as missed detections in patch-level evaluation. When the trained model was subsequently applied to the entire 3.55 km² drone mosaic together with the tiling-overlap strategy, the field-level performance increased substantially, reaching F1-Scores above 94.5% at 30–50% overlap, which represents the operational accuracy of the full pipeline rather than the model alone. However, when overlap rates became excessive, the false-detection area and processing time increased disproportionately. To address this trade-off, the proposed centroid-based tiling-movement technique maintained a 95.8% field-level detection rate with a 3.0% false-detection rate and completed the analysis within 3 min, achieving a 33% reduction in processing time compared with conventional fixed-overlap tiling. This study can contribute to establishing real-time monitoring systems in precision agriculture through drone image analysis.

Читать у издателя →

Похожие статьи