ARBA MINCH UNIVERSITY ARBA MINCH INSTITUTE OF TECHNOLOGY SCHOOL OF POST-GRADUATE STUDIES FACULTY OF COMPUTING AND SOFTWARE ENGINEERING Ethiopian Potato Tuber Disease Classification Using Deep Learning Approach

Abstract

Ethiopian potato also known as Plectranthus edulis is an important food and cash crop in the mid and high altitudes of Ethiopia, especially in the south and southwest parts of Ethiopia and it plays a major role in poverty alleviation and income generation. Its tubers are infected by bacterial disease; Gamo Chencha people call it “Kankirasho”. The structure of this disease is somehow similar to the common scab disease of the Irish potato. Rigorous analysis and investigation of the previous research indicated that no research has been conducted on the classification of Ethiopian potato tuber disease. This thesis, designed and developed a deep learning-based Ethiopian potato tuber disease classification model. The study applied both experimental and applied research designs using a mixed research approach. The images of 731 healthy (normal) and 710 infected Ethiopian potato tubers were collected from the farmland of the Gamo Chencha area. The collected images undergo different image preprocessing techniques (i.e., image cropping, image resizing, contrast stretching, and noise removal) to enhance and make the dataset ready for training and evaluation. The preprocessed datasets are partitioned into three different train-test ratios, including 70-30, 80-20, and 90-10 ratios before feeding into models. Subsequently, from these datasets split ratio 80-20 ratio obtained better performance. To handle model overfitting and enhance the model performance, this study utilized the data augmentation technique with common data augmentation operations. The experiments were conducted using python programming language with Keras API and the experimental outcomes depicted that the accuracy of the proposed convolutional neural network model, VGG16, EfficientNetB1, and MobileNetV2 were 97.23%, 83.74%, 90.31%, and 89.27% respectively. These results depicted that the proposed convolutional neural network model exceeded other models and had a great impact in reducing the classification errors imposed by the manual classification and it can increase productivity and the quality of tubers.