DETECÇÃO DE PLANTAS DANINHAS EM LAVOURAS DE MANDIOCA UTILIZANDO REDE NEURAL EM IMAGENS AÉREAS OBTIDAS POR VANT

Abstract

Cassava is a plant of great socio-economic importance in various regions of the world. However, weed infestation stands out as one of the main factors affecting its production, requiring rapid identification to mitigate the adverse effects on plantations. In the region of Iporá - GO, brachiaria grass has emerged as the predominant weed in local crops. The advent of agriculture 4.0 has introduced new technologies into agricultural practices, including drones, computer vision and machine learning. The integration of these tools has shown favorable results in the immediate identification of undesirable plants in crops. Among these methodologies, the use of convolutional neural network (CNN) models to locate objects in images has been gaining increasing strength. In particular, the YOLO (You Only Look Once) model has been recognized for its real-time detection efficiency. Evolving over iterations, the latest version, YOLOv8, was released in early 2023. However, despite the advances, YOLOv8 remains underutilized compared to its predecessors - YOLOv3, YOLOv4 and YOLOv5 - in academic studies. Therefore, the selection of YOLOv8 for this study aims to contribute scientific insights by evaluating its performance. The research uses an innovative approach to weed detection in cassava cultivation, employing the YOLOv8 model to identify cassava and weeds in aerial drone images. The results cover metric analysis, visual assessments, a publicly accessible dataset of aerial images of cassava and an assessment of the appropriate heights for this application. To validate the study's hypothesis, brachiaria grass was used as the target weed. The results highlight the promising performance of the YOLOv8 model at a height of 5 meters, reaching a mAP@0.5 of 69.7%. This metric underscores the model's effectiveness in accurately identifying weeds in images and videos captured by drones.