Banca de QUALIFICAÇÃO: ALFEU BETT MANFRIM

Uma banca de QUALIFICAÇÃO de MESTRADO foi cadastrada pelo programa.
DISCENTE : ALFEU BETT MANFRIM
DATA : 12/11/2025
HORA: 13:30
LOCAL: On line
TÍTULO:

YIELD MAPS IN PAPAYA CULTIVATION: A NEURAL NETWORK APPROACH FOR FRUIT IDENTIFICATION AND COUNTING

PALAVRAS-CHAVES:

YOLO, Artificial Intelligence, Object Detection

PÁGINAS: 35
GRANDE ÁREA: Outra
ÁREA: Multidisciplinar
RESUMO:

Brazil stands out as one of the main papaya producers, with a production of 1.138 million tons in 2023, highlighting the importance of productivity monitoring. Computer vision has proven to be promising in agriculture, already being applied in several scenarios such as disease identification and crop yield prediction. This study presents a practical application of the YOLOv11 neural network for the detection and quantification of papayas still on the tree, aiming to contribute to efficient agricultural management and the proper allocation of resources. The images used in the study were collected from an experimental plantation at UNEMAT, in Tangará da Serra – MT, where 4,008 high-resolution images (1800 × 4000 pixels) were captured on different dates and times to ensure a diversity of environmental conditions. After collection, the images were labeled using the LabelImg software, generating YOLO-compatible text files, where 94,776 objects were annotated. The YOLOv11 training was carried out on Google Colab over 100 epochs, with all images standardized to a resolution of 640 × 640 pixels. Five YOLOv11 models (n, s, m, l, and x) were trained, with the YOLOv11x model achieving the best performance, reaching 87.97% precision, 71.78% recall, 79.64% mAP50, and 52.92% mAP50-95. Videos of the plantation were also recorded and processed with a custom application designed to select and save images of the individual production of each tree. From this process, 75 new images were generated with 1,202 labeled papayas. This technique provided images with reduced background area, and when applied, the neural network model achieved an average detection rate of 97.82% of all papayas across the five YOLOv11 versions. The study demonstrated the effectiveness of the YOLOv11 network in detecting and quantifying papayas, showing the potential of this approach for productivity assessment in papaya cultivation.

MEMBROS DA BANCA:
Presidente - 138321001 - RAFAEL CESAR TIEPPO
Externa à Instituição - THAÍNA APARECIDA AZEVEDO TOSTA - UNIFESP
Externo ao Programa - 131991001 - WILLIAN KRAUSE