Deep Learning-based Classification of Faba Bean Leaf Diseases using MobileNet Trained from Scratch

Abstract

Background: Faba bean is a nutritionally rich legume crop that contributes significantly to food security and sustainable agriculture. However, its productivity is adversely affected by foliar diseases such as chocolate spot, gall and rust. Accurate and timely disease identification under natural field conditions remains challenging due to environmental variability and visual similarity among symptoms. Lightweight deep learning models provide an efficient solution for automated and real-time disease diagnosis. Methods: This study proposed a MobileNet-based convolutional neural network trained from scratch for four-class classification of faba bean leaf images (Chocolate Spot, Gall, Rust and Healthy). A dataset of 8,021 RGB images collected under field conditions was divided using an 80:10:10 stratified split for training, validation and testing. Images were resized to 224x224 pixels, normalized to the range [0,1] and augmented through random rotation, flipping and zooming to enhance generalization. The model was trained using the Adam optimizer with a learning rate of 5x10-5 for 50 epochs with early stopping. Performance was evaluated using confusion matrix analysis, precision, recall, F1-score, ROC and PR curves. Result: The proposed model achieved 97.36% training accuracy and 96.32% validation accuracy, with a final test accuracy of 95.14% on 1,605 unseen samples. ROC curves demonstrated near-perfect separability with AUC values approaching 1.00, while class-wise metrics confirmed balanced performance. These findings indicate that the lightweight MobileNet architecture can effectively support reliable and scalable faba bean disease detection under real-world agricultural conditions.