Application of deep learning in the phytosanitary detection of cocoa using computer vision

Navarro-Cabrera J.R.; Beraún-Barrantes J.G.; Cárdenas-García Á.; Lozano-Carranza C.M. · Revista Cientifica de Sistemas e Informatica (2026) · DOI: 10.51252/rcsi.v6i1.1385

Resumen

This study analyzed the application of deep learning for automated phytosanitary detection in cacao using computer vision, comparing the performance of three architectures: ResNet50, EfficientNet-B0, and Vision Transformer (ViT-B/16). A reproducible pipeline was implemented, integrating image preprocessing, five-fold stratified cross-validation, and inferential statistical analysis using repeated-measures ANOVA. The dataset consisted of 4,390 RGB images of cacao fruits distributed across three imbalanced classes: Healthy, Black Pod Rot, and Pod Borer. All models were fully fine-tuned and trained using the AdamW optimizer, early stopping, and a dynamic learning rate scheduler. The results showed mean F1-macro values above 0.96 across all architectures, with no statistically significant differences among models (F = 0.278, p = 0.7645). Training curves exhibited stable convergence and low inter-fold variability, with no evidence of overfitting. These findings indicate that system performance primarily depends on the quality of the experimental pipeline and class imbalance handling rather than on the specific architecture employed. © Los autores.

Instituciones: Universidad de Huánuco; Universidad Nacional de San Martín