UAV-based intelligent weed detection using YOLO11 and PSPNet for precision agriculture

Weeds significantly threaten rice yield and quality, necessitating precise herbicide spraying to control their growth and enhance agricultural productivity. However, conventional methods address object detection and segmentation separately, limiting their efficiency in identifying areas infested with weeds. This study introduces YOLO11-PSPNet, a combination of YOLO11-s and the Pyramid Scene Parsing Network (PSPNet), for weed detection and semantic segmentation using Unmanned Aerial Vehicle (UAV) images. A dataset was developed that comprised real-time images of rice paddy fields captured via UAV, which included weed varieties such as Echinochloa (barnyard grass), Cyperus difformis (small flower umbrella sedge), and Echinochloa colona (jungle rice). YOLO11-s detects weed-infested regions by generating bounding boxes, whereas PSPNet performs pixel-wise segmentation to ensure accurate weed localisation. Then, the proposed RAdam optimiser with a Sharpness-Aware Minimization (SAM) function was introduced to train the model YOLO11-PSPNet, which improved the training stability of the proposed model. The proposed YOLO11-PSPNet model was trained and tested on a UAV-based dataset, achieving a mAP50 of 99.56% with an inference time of 6.2ms. These results validate the efficiency of the model in precise weed detection, leading to improved crop health and higher yields. This study highlights the potential of YOLO11-PSPNet in precision agriculture for optimizing weed management using advanced deep-learning techniques.