Disease Detection in Averrhoa Carambola using image recognition techniques with deep learning models

Vargas Mahaney A.A.; Córdova Puma S.A.; Pérez Vera Y. · Journal of Agriculture and Food Research (2026) · DOI: 10.1016/j.jafr.2026.103331

Resumen

The Averrhoa carambola , a fruit with high nutritional and economic value, faces significant phytosanitary challenges, including anthracnose, early blight, and various pests that affect yield and commercial quality. This study evaluates different convolutional neural network (CNN) architectures for the automatic detection and classification of diseases and pests in carambola using the Carambola Disease Recognition Dataset (Mendeley Data), which contains 3913 images across nine classes derived from an original dataset of 559 images. Preprocessing, data augmentation, class balancing, and 5-fold stratified cross-validation were applied to compare EfficientNetB0, EfficientNetB7, EfficientNetV2-S, ResNet50, VGG16, DenseNet121, and InceptionV3. ResNet50 achieved the highest accuracy (92.8%) and F1-score (92.70%), outperforming the other evaluated architectures. These results indicate that ResNet50 performed effectively under the limited-data conditions of this study. However, the relatively small size of the original dataset may limit model robustness, and the reported performance should therefore be interpreted within this context. The proposed approach can support early and accurate disease detection, potentially contributing to reduced agrochemical use and economic losses in carambola production. Future work will focus on expanding the dataset, incorporating additional optimization strategies, and validating the models under real field conditions.

Instituciones: Universidad Nacional de San Agustín de Arequipa