A review of deep learning applications in weed detection: UAV and robotic approaches for precision agriculture

Abstract

Deep Learning (DL) has changed the face of weed detection and has greatly improved Site-Specific Weed Management (SSWM). A comprehensive review of DL-based weed detection approaches with Unmanned Aerial Vehicles (UAVs), autonomous robots, and high-resolution orthomosaic imagery is presented in this paper. Different DL models have been used in improving the accuracy of weed detection and classification in agricultural fields such as Convolutional Neural Networks (CNNs), Transfer Learning architectures, and self-supervised models. In addition, this review addresses the interoperability of DL models in automated weeding robots, real-time edge computing systems and UAV-based precision agriculture solutions, providing an integrated view of precision weed control. The review study recognizes the recent trends in detection approaches including lightweight DL networks, multimodal data fusion and UAV related developments through a systematic analysis of 90 research papers. However, the generalizability of DL models under variable environmental settings, lack of labeled datasets and limited scalability of DL techniques for large-scale agricultural purpose, still remain an issue in the field. This paper attempts to address this by critically reviewing recent advances, highlighting knowledge gaps, and suggesting future research directions to foster integration of DL in precision agriculture and efficient weed management.