Lightweight Transfer Learning for Multi-Class Cabbage Leaf Disease Detection Using MobileNetV2
Abstract
Accurate and timely recognition of leaf diseases is a prerequisite for reducing crop losses and supporting data-driven horticultural management. This study develops a lightweight transfer-learning framework for the multi-class classification of cabbage (Brassica oleracea) leaf conditions using MobileNetV2. The dataset consists of 1,600 RGB images representing eight classes: seven disease categories and one healthy class. Images were standardized through resizing to 224 x 224 pixels, min-max normalization, and training-only augmentation involving rotation, flipping, zoom transformation, and brightness perturbation. MobileNetV2 was adopted as the feature extractor because its inverted residual blocks and depthwise separable convolutions provide a favorable compromise between discriminative capacity and computational cost. The classification head was trained using the Adam optimizer, a learning rate of 1e-4, mini-batches of 32 images, and categorical cross-entropy loss. The trained model achieved 96.2% accuracy, with macro-averaged precision, recall, and F1-score of approximately 0.96. The confusion matrix indicates that most categories were separated reliably, whereas the remaining errors mainly occurred between visually similar symptoms, especially spot-like lesions and mild discoloration patterns. The model was also exported for web-oriented inference, demonstrating a feasible pathway from laboratory training to a lightweight smart-farming prototype. These results show that MobileNetV2 can provide high classification performance while preserving deployability, making it suitable as a baseline for practical plant disease screening systems in resource-constrained agricultural contexts.