MODELOS DE APRENDIZAGEM PROFUNDA PARA DETECÇÃO E SEGMENTAÇÃO DE PLANTAS DANINHAS COM IMAGENS OBTIDAS POR UAV

Abstract

Weeds compete for natural resources and interfere with the development of business plants present in the area. With the advancement of deep machine learning technology, studies have emerged proposing to detect, classify, and monitor weeds in agricultural crops for management and control. Due to the importance of the subject, this paper aimed to review the main studies applying deep learning to agricultural crops and evaluate the application of these models to detect and segment weeds in images captured by unmanned aerial vehicles (UAVs). This paper contributes to the computer vision field reviewing recent studies about the application, comparison, and performance of three models with different architectures from one another and makes public the dataset with aerial images of soybean and bean crops. The dataset was annotated, resized, and increased by the Roboflow application interface and, subsequently, trained. You Only Look Once (YOLO), Mask R-CNN with Detectron2, and U-Net were trained and evaluated by the mAP50, mAP50-90, precision, accuracy, and recall validation metrics and then compared to each other. The YOLOv8s variant achieved higher performance with a higher mAP50 value of 97%, recall of 99%, and precision of 99.7%, when compared to the other models. The results comparison of the different deep learning models in this paper showed great potential in the use of advanced deep learning algorithms to segment invasive plants.