Quantification and classification of coffee fruits with computer vision

• Autor(a) principal: Bazame, Helizani Couto
• Data de Publicação: 2021
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações da USP
• Texto Completo: https://www.teses.usp.br/teses/disponiveis/11/11152/tde-14122021-160103/
• topic: Agricultura de precisão; Cafeicultura; Coffee sector; Colheita mecanizada; Computer vision; Mechanized harvesting; Precision Agriculture; Visão computacional; YOLO; YOLO

Coffee is one of the most consumed and traded beverages in the world. Knowledge about the yield and maturation stage of coffee fruits before and after the harvest is still a challenge for the coffee sector. The development of a system that allows obtaining this information quickly and non-invasively is essential for the efficient management of the crop. Advances in monitoring the coffee crop should allow for the generation of maps that present essential information for diagnosing the spatial and temporal variability of the crop and, consequently, for the efficient use of precision agriculture techniques. One of the alternatives used to estimate the yield and ripening stage of coffee fruits would be the use of computer vision techniques based on object detection and classification. The use of computer vision offers a low-cost and accessible solution, with great potential for improving the monitoring of coffee plantations. This study was divided into three chapters that present the use of computer vision models based on the YOLO neural network architecture to detect coffee fruits under different contexts. In chapter 1, the model is used to detect and classify coffee fruits on tree branches, a tool that can help small and large producers to objectively decide when to start the harvest. In chapter 2, the model is used to detect and count coffee fruits during mechanized harvesting, which allows the generation of yield maps for the harvested areas. In chapter 3, the model is used to detect and classify coffee fruits at different stages of maturation during mechanized harvesting, which allows for the spatialization of the coffee maturation stage for the harvested areas. The computer vision models based on the YOLOv4 architecture and an input image with a resolution of 800x800 pixels had mean average precision (mAP) of 81.2%, 83.5% and 91.8% for the scenarios experienced in chapters 1, 2 and 3, respectively. The yield map estimated from the detections obtained by the model was able to explain 81% of the variance of the yield map used as reference. The knowledge of the spatial and temporal variability of information such as productivity and maturation stage are essential for the implementation of precision agriculture techniques in coffee crops.