Edge-deployable segmentation and prescription mapping of post-emergence weeds in sugar beet crops for UAV-based precision spraying

Weeds significantly challenge sugar beet cultivation by reducing yields and increasing herbicide usage. Conventional broadcast spraying inflates production costs and raises environmental concerns due to chemical overuse. To address these challenges, this study presents a comprehensive deep learning (DL)-based framework for site-specific weed detection and precision herbicide application using UAV-acquired imagery. High-resolution RGB ortho-mosaic data from experimental sugar beet fields were used to train and evaluate multiple semantic segmentation (U-Net, PSPNet, DeepLabv3) and instance segmentation (YOLOv8, Mask R-CNN) models. U-Net, coupled with a ResNet-34 backbone, achieved the highest segmentation accuracy, with IoUs of 0.85 for Sugar beet and 0.72 for weeds. Prescription maps derived from segmented weed cover suggested a theoretical herbicide savings rate of up to 82.88 % compared to conventional uniform spraying. Instance segmentation was also performed using YOLOv8, YOLOv9, YOLO11, and Mask R-CNN to detect weed patches within the crop canopy. YOLOv8 outperformed Mask R-CNN in instance segmentation mAP (0.728 vs. 0.627), while Mask R-CNN achieved higher classification precision. Edge deployment was explored using optimized and quantized YOLOv8 variants, with the YOLOv8-SB model offering real-time inference speeds exceeding 48 FPS. This research demonstrates a scalable and field-deployable approach for weed mapping and precision spraying in dense weed patches, significantly reducing chemical use while enhancing site-specific weed management practices. • A deep learning approach for aerial weed segmentation and localization during the post-emergence growth stage. • Mapping of weeds based on DL segmentation models for real-time weed spraying and removal. • Identifying areas with high weed density is crucial for optimizing herbicide usage.