Hierarchical Convolutional Neural Networks for Leaf Disease Detection

The presence of plant leaf diseases poses a threat to food security and their detection becomes increasingly challenging when multiple types of diseases affecting various plant species are present. In such scenarios, the high number of classes necessitates careful consideration of the choice of model, as a single Convolutional Neural Network (CNN) model may have a higher likelihood of overfitting. This paper advocates for the use of a hierarchical approach, incorporating multiple CNN models, as an alternative to using only one model, in order to address the problem of unbalanced data. The effectiveness of the proposed method is demonstrated using a public dataset of 87,000 images of diseased and healthy leaves of 14 different plant species categorized into 38 classes. The proposed method achieves an accuracy of 97.17% on a held-out test set, demonstrating the generability of the hierarchical models over one single model.