Malloy-dan/maize-disease-system

Ghana Maize Disease Detection System - offline-capable app for smallholder farmers. MobileNetV2 transfer learning + TFLite + React Native. # Offline-Capable Maize Disease Detection System A deep learning-powered mobile application for the early detection of maize diseases, designed specifically for smallholder farmers in Ghana's northern sector. The system runs **fully offline** on low-cost Android devices (2 GB RAM, Android 6.0+), making it accessible to farmers who lack reliable internet connectivity. This project is the artefact of a final-year project report at C. K. Tedam University of Technology and Applied Sciences. It uses MobileNetV2 transfer learning, converts the trained model to a lightweight TensorFlow Lite file, and embeds it in a React Native app with English and Twi language support. --- ## Background & Motivation Maize (*Zea mays*) is Ghana's most important cereal crop and the primary staple food for over 31 million Ghanaians. Smallholder farmers in the northern sector — particularly the Upper East Region — rely heavily on maize for food security and household income. However, they face severe challenges in disease management: - Ghana's **extension officer-to-farmer ratio** is about **1:1,500**, three times worse than the recommended 1:500. Rural farmers have minimal access to expert diagnosis. - Visual symptoms often appear only after a disease has progressed, and maize disease outbreaks can cause **yield losses of 30–80%**. - Most smallholder farmers cannot distinguish between diseases with similar symptoms, leading to incorrect or absent treatment. - Most existing mobile disease-detection apps **require internet access**, which is unreliable in rural northern Ghana. - Most models are trained only on laboratory images and **foreign disease sets**, and do not cover African diseases such as **Maize Streak Virus (MSV)** and **Maize Lethal Necrosis (MLN)**, nor provide local-language support or a rejection class for non-maize images. This system addresses these gaps by providing an accurate, lightweight, fully offline tool with English/Twi support, a `Not_Maize` rejection class, a …