Automatic diagnostic system for mango (Mangifera indica L.) leaf diseases trees based on MobileNetV2

Lozada C.R.S.; Flores E.A.; Cayllahua E.C.; Ramos L.J.L. · Acta Agronomica (2026) · DOI: 10.15446/acag.v74n1.123188

Resumen

Mango cultivation (Mangifera indica L.) is vulnerable to multiple leaf diseases that severely affect productivity, making early identification essential to mitigate economic losses. The objective of this research was to train and evaluate a model for classifying mango leaf diseases based on the MobileNetV2 architecture using transfer learning. A dataset of 4000 images was used, with a stratified partition of 70 % for training, 15 % for validation, and 15 % for independent testing (n= 600). A two-phase training strategy was implemented in Python (TensorFlow/Keras), and model performance was evaluated using multidimensional metrics such as Cohen’s Kappa coefficient, MCC, ROC curves, and an analysis of predictions confidence. The model demonstrated a high capacity for generalisation on the independent test set, achieving an accuracy of 87.67 %, a Kappa value of 0.859, and an MCC of 0.862. The ROC analysis revealed a micro-average AUC of 0.9926. The main diagnostic difficulty was distinguishing early symptoms of Gall Midge from healthy leaves (22.67 % error). The results confirm that confidence analysis (mean= 0.855 for correct predictions vs. mean = 0.580 for errors) enables the establishment of diagnostic confidence thresholds. It is concluded that MobileNetV2 provides an optimal balance between computational efficiency and diagnostic reliability, representing a viable solution for phytosanitary monitoring in precision agriculture. © (2026), (Universidad Nacional de Colombia). All rights reserved.

Instituciones: Universidad Nacional Agraria La Molina