A Comprehensive Review of Deep Learning Approaches for Automated Farm Weed Classification

Abstract

One of the biggest obstacles to agricultural output is weed infestation, which causes a significant drop in crop production, increases production costs, and affects food security. Conventional ways of controlling weeds, such as hand weeding and using herbicides, are usually very laborious, expensive, and unsustainable. In recent years, deep learning has become an attractive option in automated weed detection and classification in precision agriculture. This paper is a review of the extensive classification of farm weeds through deep learning models with special emphasis on object detection networks, convolutional neural networks (CNNs), transfer learning models, transformer-based models and data augmentation techniques. Applicable literature was methodically evaluated with reference to datasets, model structures, and performance parameters, including accuracy, F1-score, and mean average precision (mAP). The review indicates that other models, such as YOLO variants, ResNet, EfficientNet, and Vision Transformers, have demonstrated high classification accuracy under controlled conditions. Nevertheless, issues such as limited dataset diversity, inadequate real-world generalisation, excessive computational complexity, and standardised evaluation schemes continue to impede at-scale implementation. The research results have concluded that, in the future, lightweight models, larger-scale, more diverse data, and the integration of deep learning with IoT and autonomous systems should be the subject of research to make the processes of monitoring and controlling weeds in modern agriculture fully automated and sustainable.