Precision weed detection using UAVs and deep learning: Models, paradigms, and challenges

Weeds continue to pose a major challenge to agricultural productivity, contributing to an estimated 31.5 % yield reduction and a global economic loss of $32 billion. Recent advancements in artificial intelligence, particularly deep learning (DL), have enabled the development of precise and automated weed detection systems. This review explores the state-of-the-art in deep learning-based weed detection, with a focus on unmanned aerial vehicles (UAVs), imaging sensors, and computer vision techniques. UAVs, equipped with various imaging sensors, offer high-resolution imagery for site-specific weed monitoring. DL models including one-stage detectors such as YOLO, RetinaNet, EfficientDet, and CenterNet, and two-stage detectors like Faster R-CNN and Mask R-CNN have demonstrated strong performance in weed localization and classification tasks. While one-stage models provide faster inference for real-time applications, two-stage models typically offer higher accuracy at the cost of increased computational demands. Segmentation models such as U-Net and DeepLab further enable pixel-level weed mapping, crucial for estimating coverage and guiding targeted interventions. This paper also outlines the complete training pipeline for DL models, including data acquisition, annotation, preprocessing, model selection, and evaluation using metrics like precision, recall, mAP, and inference time. Additionally, the role of supervised, semi-supervised, and self-supervised learning paradigms is discussed in addressing data scarcity and annotation challenges. Future research should focus on improving model generalization, enabling edge deployment for real-time field applications, and developing robust, diverse datasets. These advancements will facilitate the scalable adoption of precision weed detection systems, supporting more sustainable and efficient agricultural practices.