IS http://repository.mtu.edu.et/xmlui/handle/123456789/12 2026-05-27T19:33:17Z 2026-05-27T19:33:17Z Adugnaw, Demeke Andualem, Amogne Gardie, Birhanu Ashebir, Desalegn Aytenfsu, Misganaw http://repository.mtu.edu.et/xmlui/handle/123456789/238 2025-02-28T08:13:27Z 2023-06-15T00:00:00Z

Automatic Enset Plant Disease Identification Using Deep Convolutional Neural Network Adugnaw, Demeke; Andualem, Amogne; Gardie, Birhanu; Ashebir, Desalegn; Aytenfsu, Misganaw Most Ethiopian people use it as a main food by preparing bread from its stems and roots by using traditional fermenting techniques. Enset plant is a multi-objective traditional crop cultivated and widely used in the south and southwestern part of Ethiopia. Generally, the Enset plant has a predominant role in people living in the southern part of Ethiopia. The Enset plant is vulnerable to different diseases like Bacterial wilt, Sigatoka, and others. This leads to enforce farmers to replace Enset with other crops. These diseases have a significant impact on Enset production. Even though farmers can stop and prevent such diseases at an early stage, they are unable to determine which particular disease affects the crop. Based on our survey, not only farmers but also some pathologists cannot easily identify which disease affects the crop. Therefore, to address this issue, we created an AI model that automatically recognizes and detects the diseases that affect the plant. This model aims to support farmers and pathologists to recognize and detect these diseases in the early stage. We have collected infected and normal image datasets from Sheka, Bench-Sheko, and Keffa districts of Enset cultivation fields. Our model includes various computer vision techniques such as image preprocessing, segmentation, feature extraction, and classification. As a comparison, we have used different convolutional neural network-based models such as Alex Net, VGG Net, and Res Net. A ResNet model with SoftMax classifier achieved 76% testing accuracy. Alex Net CNN-based model achieved 90% testing accuracy. The other CNN-based pre-trained model is VGG Net which achieved 94% testing accuracy. Even though the pre-trained models have promising results, they need to maximize the prediction accuracy and minimize lose value.So, we have built a new CNN-based architecture to control overfitting and increase classification accuracy and it achieves 98.2%

2023-06-15T00:00:00Z