Efficient Mulberry Leaf Disease Detection in Bangladesh: A Lightweight Approach for Real-Time Applications

Bombyx mori, the silkworm species that feeds on mulberry leaves, produces the most prevalent variant of silk worldwide. Unfortunately, mulberry trees suffer from leaf diseases that spread quickly and cause significant losses in production. Identifying diseases by experts on large scale is time-consuming and often inaccessible for rural people. Despite recent advances in deep learning and image processing, State-of-the-art models tend to be computationally expensive and must be used online. This study makes use of a regional dataset that comprises of mulberry leaf images captured from Rajshahi, the principal sericulture region in Bangladesh, to align with the environmental factors of the country. Firstly, we present an extremely efficient and lightweight approach enhanching a pre-trained MobileNetV2 model to classify among three categories of leaf photos: disease free, leaf spot, and leaf rust. Upon comparison with base MobileNetV2 and existing works on this dataset, this model exhibited efficacy in accuracy, recall, precision, and F1-score, attaining remarkable values of 98.63%, 98.53%, 98.63%, and 98.63%, respectively after 5-fold cross-validation. Then, we performed an explainability analysis utilizing Grad-CAM visualization to validate the capability of our suggested method. We also assessed the impact of preprocessing techniques on classification outcomes. The proposed architecture, having only 145K parameters, remains surprisingly lightweight and outperforms the base MobileNetV2 model in our experiments. According to the results, our suggested approach correctly identifies mulberry leaf diseases, supporting sericulture growth in rural areas.