Annotation-efficient weed mapping in sorghum fields using two-stage U-Net crop segmentation and HSV greenness analysis

Accurate weed mapping is essential for site-specific weed management in sorghum, but pixel-level weed annotation is labor-intensive and limits practical deployment. This work proposes an annotation-efficient two-stage approach using Unmanned Aerial Vehicle (UAV) RGB imagery that first segments crop pixels with U-Net and then detects weeds in non-crop regions using greenness-based analysis. Convolutional Neural Network (CNN)- and transformer-based encoders were evaluated for crop segmentation, and hue, saturation, value (HSV) filtering was compared with the Excess Green index (ExG) and the Color Index of Vegetation Extraction (CIVE) for weed detection. Among the evaluated configurations, the best-performing combination used U-Net with an EfficientNet-B5 encoder for crop segmentation and HSV-based filtering for weed detection. The approach was evaluated across sorghum growth stages (BBCH 15, 17, and 19) and produced practical weed-mapping outputs while substantially reducing annotation requirements, because only crop masks were needed and manual weed labeling was avoided. In total, 4,470 weed instances did not require manual annotation. When compared with a fully supervised end-to-end multi-class model, the proposed two-stage framework showed lower but practically useful segmentation performance while substantially reducing annotation requirements. The method also generated field-scale weed distribution maps and weed-cover estimates that can support targeted scouting and site-specific herbicide application. These results support the proposed approach as a practical and scalable alternative for precision weed management under real field conditions.