A Deep Learning Approach to Multi-class and Multi-label Cassava Leaf Disease Detection Using MobileNetV2 and Image Augmentation

This paper presents the development of a deep learning model using pre-trained models to detect multiple cassava leaf diseases using simultaneous multi-class and multi-label classification. Cassava diseases often occur as single or mixed infections, complicating visual diagnosis critical to food security in tropical regions.Existing models, designed primarily for single-label classification, struggle with overlapping symptoms. To address this, a dataset of 10,000 expert- annotated cassava leaf images was compiled from PlantVillage. The MobileNetV2 model, optimized for mobile deployment, was trained using binary cross-entropy loss with Sigmoid activation for Multi-label classification and Categorical cross-entropy loss with Softmax activation. The model achieved 95% accuracy, outperforming NasNet, demonstrating effective real-time, in-field disease diagnosis. This research advances agricultural AI applications by enabling scalable, mobile-compatible cassava disease detection.