Deep learning approaches for automated cassava leaf disease detection.

Abstract

Cassava is an important food security crop in Uganda, but its production is significantly affected by diseases such as Cassava Mosaic Disease (CMD) and Cassava Brown Streak Disease (CBSD). Early and accurate identification of these diseases remains challenging due to reliance on manual visual diagnosis, which is subjective, time-consuming, and constrained by limited access to agricultural extension services. This study aimed to develop, evaluate, and deploy deep learning models for automated cassava leaf disease detection using the Makerere University Cassava Image Dataset. An experimental research design was adopted, involving the development and comparison of a baseline Convolutional Neural Network (CNN) trained from scratch and a MobileNetV2 transfer learning model. The dataset was cleaned by removing corrupted images and balanced by selecting 1,430 images from each of three classes: Healthy, CMD, and CBSD, resulting in a total of 4,290 images. The data were partitioned into training, validation, and testing sets, and model performance was evaluated using accuracy, precision, recall, F1-score, and confusion matrix analysis. The baseline CNN achieved a test accuracy of 54.20% and exhibited substantial overfitting, indicating limited generalization capability. In contrast, the fine-tuned MobileNetV2 model achieved an overall test accuracy of 93%, with precision, recall, and F1-score of 0.93 across the three classes. The MobileNetV2 model also demonstrated substantially reduced misclassification and better generalization than the baseline CNN. The best-performing MobileNetV2 model was subsequently deployed as a Streamlit-based prototype capable of providing real-time cassava leaf disease classification. The findings demonstrate that transfer learning provides a more effective and computationally efficient approach than training a CNN from scratch for automated cassava disease detection. The developed prototype provides a foundation for AI-based agricultural decision-support systems that can support timely disease diagnosis, improve disease management, and contribute to cassava productivity and food security in Uganda. However, further field validation using independent datasets and testing with farmers and agricultural extension workers are recommended before large-scale deployment.