Advances on deep learning for proximal image-based weed recognition and control under authentic farmlands: A state-of-the-art review

This review systematically examines deep learning (DL)-based frameworks applied to weed recognition using proximal imagery under authentic farmland conditions. It explores advances in classification, detection, and segmentation networks by analyzing 349 relevant works published between 2015 and 2025. Special attention is given to the challenges of training DL models with datasets that reflect real-field variability, such as diversity in weed species, growth stages, densities, and lighting conditions, which hinder the models' generalization and real-world applicability. The review also compiles 34 publicly available datasets acquired in open-field crop conditions, as well as common implementation steps and strategies adopted by researchers to develop robust models. From the retrieved classification works, the ResNet and VGG network families are most frequently implemented, while for detection, the YOLO family appears in 53.48% of the works. UNet is the most common choice for segmentation. Particular focus is also placed on studies that have deployed DL-based vision systems on smart weeders and evaluated them in authentic farmland environments; 26 such works were identified, and key variables that affect model performance in real-time weed detection and control were extracted, including the evaluation metrics used to assess intelligent weeders' effectiveness. Finally, this work outlines future directions for improving the robustness of DL models, including enhancing dataset diversity and other alternatives for achieving a more effective DL-based weed recognition system. • A comprehensive review of deep learning models for weed recognition in real-world farmlands. • A collection of publicly available crop–weed datasets. • Deep learning models deployed on smart weeders and evaluated in actual crop fields. • Common metrics used to evaluate smart weeders. • Discussion of previous works and future recommendations for the research community.