Development of Cassava Leaf Disease Detection System using Convolutional Neural Network-Based Zebra Optimization Algorithm

Abstract

Cassava is a widely cultivated root crop valued for its starchy tubers and nutritional importance in tropical regions; however, its productivity is severely affected by diseases and pests, resulting in substantial yield losses for farmers. Although recent advances in deep learning offer effective solutions for automated disease detection, challenges related to model accuracy and hyperparameter tuning remain. This study optimizes a Convolutional Neural Network (CNN) using the Zebra Optimization Algorithm (ZOA) for cassava leaf disease detection. A total of 19,620 cassava leaf images were obtained from Kaggle and categorized into four classes: Cassava Mosaic Disease, Cassava Green Mite, Cassava Bacterial Blight, and healthy leaves. Image preprocessing techniques, including cropping, grayscale conversion, and normalization, were applied to enhance training efficiency. The ZOA was employed to optimize key CNN hyperparameters, specifically the number of neurons and dropout rate. The ZOA-CNN model was implemented using MATLAB R2023a and evaluated using false positive rate, specificity, sensitivity, accuracy, and recognition time. Experimental results show that the ZOA-CNN achieved an FPR of 1.11%, specificity of 98.89%, sensitivity of 95.28%, accuracy of 98.12%, and recognition time of 37.70 s, outperforming the conventional CNN. These results demonstrate that the ZOA-CNN improves detection accuracy, reduces false detections, and enhances computational efficiency, making it suitable for real-world cassava disease monitoring applications.