Publicação
Deep learning applied to detection and segmentation of european canker on apple tree branches
| Resumo: | European canker, caused by Neonectria ditissima, is one of the main diseases affecting apple orchards and can compromise both plant health and fruit production. The visual identification of canker lesions is usually performed manually, which makes the process dependent on specialist observation and limits its scalability under field conditions. In this context, this work investigates the use of deep learning models for the automatic detection and segmentation of European canker in apple tree branches. The study used an image dataset of healthy and infected branches, annotated for the branch and canker classes, and a synthetic orchard-like dataset generated by compositing annotated branch instances over frames extracted from orchard videos. The evaluated methodology included preliminary experiments with YOLOv8 and YOLOv12 variants, in which the YOLOv8 family showed more stable performance for this dataset, followed by the final assessment of selected detection and segmentation strategies. For detection, yolov8l was selected and achieved strong test-set performance, although the class-specific results showed that canker localization remained more challenging than branch localization. For segmentation, direct yolov8x-seg, yolov8l boxes combined with SAM 2, and Mask2Former were compared under a common validation procedure. Mask2Former obtained the best overall segmentation overlap, with Intersection over Union of 0.751 and Dice of 0.858, while direct yolov8x-seg achieved the best canker-specific delineation, with canker Intersection over Union of 0.772 and canker Dice of 0.871. The prompt-based SAM 2 pipeline produced high precision but did not recover lesion contours as effectively as direct supervised segmentation. Additional qualitative video outputs and an auxiliary inference-time benchmark were used to inspect the behavior of the models on synthetic orchard video sequences. Overall, the results indicate that deep learning can support the detection and segmentation of European canker in apple tree branches, with direct yolov8x-seg representing the most suitable strategy among the evaluated alternatives when the main objective is accurate canker delineation. |
|---|---|
| Autores principais: | Facundo, Bruno Raphael |
| Assunto: | European canker Apple tree branches Deep learning Object detection Image segmentation |
| Ano: | 2026 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Politécnica de Bragança |
| Idioma: | inglês |
| Origem: | Biblioteca Digital da UPB |
| Resumo: | European canker, caused by Neonectria ditissima, is one of the main diseases affecting apple orchards and can compromise both plant health and fruit production. The visual identification of canker lesions is usually performed manually, which makes the process dependent on specialist observation and limits its scalability under field conditions. In this context, this work investigates the use of deep learning models for the automatic detection and segmentation of European canker in apple tree branches. The study used an image dataset of healthy and infected branches, annotated for the branch and canker classes, and a synthetic orchard-like dataset generated by compositing annotated branch instances over frames extracted from orchard videos. The evaluated methodology included preliminary experiments with YOLOv8 and YOLOv12 variants, in which the YOLOv8 family showed more stable performance for this dataset, followed by the final assessment of selected detection and segmentation strategies. For detection, yolov8l was selected and achieved strong test-set performance, although the class-specific results showed that canker localization remained more challenging than branch localization. For segmentation, direct yolov8x-seg, yolov8l boxes combined with SAM 2, and Mask2Former were compared under a common validation procedure. Mask2Former obtained the best overall segmentation overlap, with Intersection over Union of 0.751 and Dice of 0.858, while direct yolov8x-seg achieved the best canker-specific delineation, with canker Intersection over Union of 0.772 and canker Dice of 0.871. The prompt-based SAM 2 pipeline produced high precision but did not recover lesion contours as effectively as direct supervised segmentation. Additional qualitative video outputs and an auxiliary inference-time benchmark were used to inspect the behavior of the models on synthetic orchard video sequences. Overall, the results indicate that deep learning can support the detection and segmentation of European canker in apple tree branches, with direct yolov8x-seg representing the most suitable strategy among the evaluated alternatives when the main objective is accurate canker delineation. |
|---|
Atividades financiadas
Não foram encontrados projetos financiados.