Evaluation of state-of-the-art models for improving the diagnosis of mango tree diseases through supervised learning of symptom images in Burkina Faso

Abstract In Burkina Faso, the mango industry faces numerous phytosanitary constraints. This situation is exacerbated by producers’ lack of knowledge about diseases and their symptoms and the low uptake of control technologies. This study aims to develop a tool using artificial intelligence-based approaches for better integrated management of the main diseases affecting mango trees. Surveys conducted in 40 mango orchards yielded 11,001 images of the characteristic symptoms of three major diseases, including 4,527 for anthracnose caused by Colletotrichum gloeosporioides , 3,038 for bacterial disease caused by Xanthomonas citri pv. mangiferaeindicae and 3,436 for dieback caused by Lasiodiplodia spp. The data underwent several preprocessing steps. To ensure a balance between disease classes, 3,000 images per class were selected, for a total of 9,000 images. The images were then annotated and used for training with the pre-trained YOLOv11 model. Following training, three models were formed : V2-3 M (MAP@50 = 50,3%; precision = 55,1%; recall = 49,1%; F1-score = 51,9%), V3-3 M (MAP@50 = 23%; precision = 56,8% ; recall = 26,7%; F1-score = 36,3%), and V4-3 M (MAP@50 = 46,8%; precision = 52%; recall = 45,3%; F1-score = 48,4%). The most effective model is the V2-3 M, as it combines 55,1% accuracy, 50,3% (MAP@50, 49,1% recall and F1-score = 51,9%). In order to improve diagnosis and phytosanitary management in orchards, it would be useful to integrate treatment recommendations into the model, then deploy it on a mobile application and promote it to producers. Future work will involve dataset expansion, comparative evaluation with alternative deep learning models, and cross-regional validation to improve model robustness and generalizability.