Implementation of a Convolutional Neural Network for Automated Cassava Disease Detection on a Raspberry Pi

Cassava is a staple food crop for millions of smallholder farmers in South East and South South Nigeria, and fungal and viral diseases including cassava mosaic virus and cassava brown streak disease cause yield losses of up to 70 percent when undetected. This project implemented a convolutional neural network-based cassava leaf disease classifier on a Raspberry Pi 4, designed as a portable field diagnostic tool accessible to extension workers without internet connectivity. A MobileNetV2 model was trained on the Cassava Leaf Disease Dataset (21,367 images, five classes) using TensorFlow on a GPU workstation. The trained model was converted to TensorFlow Lite format for deployment on the Raspberry Pi. An 8MP camera module captured leaf images, and the inference script ran on the device in less than 1.4 seconds. Classification accuracy on the held-out test set was 89.7 percent (top-1 accuracy). The confusion matrix showed the highest misclassification between cassava mosaic disease and cassava bacterial blight at 8.3 percent error, attributed to visual similarity between these two conditions. The TensorFlow Lite model size was 8.2 MB, suitable for offline embedded deployment. The prototype was demonstrated to 12 agricultural extension workers in Imo State, who rated ease of use at 4.2 out of 5. The study recommends collecting a locally curated dataset of cassava leaf images from South East Nigerian farms to improve model accuracy for regional disease variants and creating a Igbo-language voice output feature for accessibility.

Keywords: convolutional neural network, cassava disease detection, Raspberry Pi, TensorFlow Lite, South East Nigeria