FOR THE PARTIAL FULFILLMENT OF THE REQUIREMENT FOR THE AWARD OF THE DEGREE OF MASTER OF SCIENCE IN INFORMATION TECHNOLOGY

Abstract

Ethiopia is one of the world's developing countries, and its economy depends on both agriculture and industry. Among agricultural products, the banana crop is one of the most important foods for both humans and animals. Banana diseases are natural factors that can severely affect banana crops, ultimately reducing product quality and reducing yield. Manually, through human vision, categorization, classification, and management of banana leaf diseases is challenging work because the similarity of disease symptoms makes different farmers and domain experts diagnose the same diseases as different diseases, and vice versa. Previously conducted research mainly focused on tasks of identification. So, to overcome the problem, this research work was conducted. The main objective of this work is to develop a banana plant leaf disease classification model using deep learning. From different banana diseases, this study primarily focuses on common diseases such as black sigatoka, yellow sigatoka, and Cordina leaf spot that affect the leaves of banana plants. In order to conduct an experiment, researchers collected and preprocessed a total of 4020 banana leaf images from four different classes. This study followed an experimental research design with a mixed research approach by incorporating both quantitative and qualitative approaches, for collecting and analyzing the existing state of banana diseases through the Python programming language for model development. Researchers trained and conducted hyper-parameter tuning through a grid search algorithm. In this work, researchers selected four algorithms, such as CNN, VGG19, EfficientNetB2, and ResNet50, based on a deep study conducted of previous related work to identify the best-performing classifier. The experimentation was conducted using selected models through a data split of 70/20/10 based on different evaluation metrics and evaluated models’ classification performance. The last output shows that the EfficientNetB2 model outperforms the remaining models through test accuracy of 0.99, 0.99, and 0.99 in training, validation, and testing, respectively. At the end, depending on the results obtained, a proposed model prototype was developed by researchers to classify banana leaf diseases