LIGHTWEIGHT MOBILE NEURAL NETWORK FOR REAL-TIME BANANA LEAF DISEASE DETECTION USING EDGE AI

Abstract This paper addresses the challenge of detecting fungal and bacterial diseases in banana plants, which heavily impact yield and quality. While deep learning techniques have proven effective for disease detection with high accuracy, they often demand substantial computational resources. This makes many existing models impractical for use in rural farming areas where access to powerful hardware is limited. To overcome these limitations, the paper presents Mobile-BananaNet, a lightweight deep learning model specifically designed for real-time banana disease detection on low-resource mobile devices. The architecture leverages depth wise separable convolutions combined with a carefully optimized feature extractor to significantly reduce computational load without sacrificing accuracy. This design ensures the model runs quickly and efficiently on typical mid-range smartphones or edge devices. We conducted extensive experiments using a balanced dataset consisting of four major banana diseases. Mobile-BananaNet achieved an impressive classification accuracy of 97.8%, demonstrating its effectiveness in identifying disease conditions accurately. Additionally, the model size is only 5.3 MB, which is considerably smaller than many conventional deep learning models, permitting easy deployment on mobile platforms. The inference time of 42 milliseconds per image highlights its capability for real-time detection, providing immediate feedback to farmers. Overall, Mobile-BananaNet offers a practical and scalable approach for early disease diagnosis in banana cultivation. It enables farmers in rural and resource-constrained environments to use affordable mobile devices for on-site monitoring, ultimately promoting timely treatment and better crop management. This edge AI solution bridges the gap between advanced machine learning and the needs of traditional agriculture.