Weed detection based on deep learning from UAV imagery: A review

Weeds are undesirable plants that compete with crops for essential resources such as light, soil, water, and nutrients. Additionally, they can harbor pests that reduce crop yields. In traditional agriculture, weed control is based on applying pesticides throughout the agricultural field, resulting in soil damage, environmental contamination, damage to farm products, and risks to human health. Precision agriculture (PA) has evolved in recent years thanks to sensors, hardware, software, and innovations in unmanned aerial vehicle (UAV) systems. These systems aim to improve the localized application of chemicals in weed control by using advanced image analysis techniques, computer vision, deep learning (DL), and geo-positioning (GPS) to detect and recognize weeds. This subsequently facilitates the implementation of specific control mechanisms in real environments. Recently, automatic weed detection techniques have been developed using UAV imagery. However, these face a significant challenge due to the morphological similarities between weeds and crops, such as color, shape, and texture, which makes their practical and effective differentiation and implementation difficult. This paper presents a systematic literature review (SLR) based on 77 recent and relevant studies on weed detection and classification in UAV imagery using DL architectures. The analysis focuses on key aspects such as using UAVs and sensors, image acquisition and processing, DL architecture, and evaluation metrics. The review covers publications from 2017 to June 2024 from WoS, Scopus, ScienceDirect, SpringerLink, and IEEE Xplore databases. The results allowed the identification of various limitations, trends, gaps, and opportunities for future research. In general, there is a predominant use of multirotor UAVs, particularly the DJI Phantom with RGB sensors, showing a trend towards the integration of multiple sensors (multispectral, LiDAR) operating at heights of around 10 meters, providing good spatial coverage in data acquisition. Likewise, the rapid development of deep learning architectures has driven CNN models such as ResNet for classification, YOLO for detection, U-Net for semantic segmentation, and Mask R-CNN for weed instance segmentation, with a tendency towards new Transformer-based and hybrid architectures. The most common metrics used to evaluate these models include precision, recall, F1-Score, and mAP.