Automatic detection of coffee leaf diseases through pattern recognition

Abstract

Early detection of coffee leaf diseases is essential to reduce production losses; however, visual field diagnosis presents limitations associated with subjectivity and environmental variability. The objective of this study was to design and evaluate a hybrid pattern-recognition model to classify healthy coffee leaves and those affected by coffee leaf rust and brown eye spot using images captured under real field conditions in Saposoa (San Martín, Peru). A proprietary dataset of 1,500 images validated by a specialist (500 per class) was used and expanded through controlled data augmentation to 6,000 balanced images. ResNet18 was employed as a feature extractor using transfer learning, and three supervised classifiers were compared: SVM, Random Forest, and XGBoost. Model performance was evaluated using 10-fold stratified cross-validation and an independent test set (20%). The ResNet18 + SVM model achieved the best performance, with an accuracy of 0.9742, a macro F1-score of 0.9730, and a macro-AUC of 0.9968, outperforming Random Forest (accuracy = 0.9367) and XGBoost (accuracy = 0.9583). Inferential analysis using ANOVA and Tukey’s HSD test confirmed statistically significant differences among models (p < 0.001). The results demonstrate the robustness and feasibility of the proposed approach to support coffee phytosanitary diagnosis under real-world field conditions.