Machine Learning and Deep Learning Techniques for Weed Detection in Agricultural Fields: Recent Advances, Challenges, and Future Directions

Abstract

Weed infestation is a key biotic limitation on agricultural productivity. Weeds compete with crops for light, water, nutrients and space, while increasing production costs and, in many instances, functioning as reservoirs for pests. Traditional weed management is mainly dependent on manual, mechanical and blanket chemical control. Herbicides are effective; yet, indiscriminate use increases input cost, accelerates herbicide resistance development, and raises environment and food safety problems. Precision agriculture so increasingly relies on computer vision, machine learning (ML), deep learning (DL), unmanned aerial vehicles (UAVs) and agricultural robotics to detect weeds and to assist site-specific treatment. This critical review summarizes the evolution of handcrafted color, texture, shape, spatial and spectral features with support vector machines, random forests, k-nearest neighbors and artificial neural networks to modern convolutional neural networks, YOLO-family object detectors, encoder-decoder segmentation models, transformers and hybrid architectures. The entire operational pipeline is investigated, including sensing, picture pre-processing, augmentation, feature representation, classification, detection, segmentation, model evaluation, edge deployment, and closed-loop actuation. Recent literature demonstrates that in the real-time object identification, YOLO versions are leading the way while U-Net and other networks are still dominating in the semantic segmentation and ResNet / VGG family are commonly employed in classification. However, the outstanding performance in the datasets often deteriorates in the real field settings because of the illumination variations, occlusion, crop-weed morphological resemblance, small target size, growth stage variations, class imbalance, motion blur, and geographic or seasonal domain shift. High-priority research directions include cross-domain generalization, diversity of datasets, external validation, uncertainty-aware prediction, open-set recognition, multimodal sensing, self-supervised learning, domain adaptation, lightweight edge inference and integration with robotic or variable-rate weed control. The primary takeaway is that future advancement must go beyond incremental gains in benchmark accuracy toward robust, safe, energy-efficient, economically feasible and agronomically verified intelligent weed-management systems.