Developing Common Bacteria Blight Disease Detection Model for Common Bean Using Deep Learning Approach

Abstract

Abstract In many farm locations in Ethiopia, common bacterial blight (CBB), one of the common bean's known diseases, impacts the crop, lowers yield by up to 45 percent of the production, and also has an impact on seed quality. Recently, many researchers have tried to overcome this disease by conducting field surveys, crop management activities, crop rotation, chemical treatment, proper cultural practices, and integrated disease management. However, this method is a very tedious, time-consuming, and costly technique that requires more experts. To overcome these issues, a modern deep learning approach is proposed for the early detection of common bean disease. The three main phases are followed for the proposed approach. Firstly, the collection of healthy and diseased common bean images with the help of domain experts from different agricultural research centers is done. Then the design of a modern convolutional neural network that can detect and classify the input image as diseased or healthy is done. Lastly, the designed model is trained and evaluated. During the classification, the designed model is assessed against the performance of two pre-trained models (VGG16 and InceptionV3) to accurately detect the common bean disease. The performance of the proposed model is evaluated using a dataset that contains total images of 3,135 with two classes of healthy and common bacterial blight disease. We have used data augmentation techniques to generate more images to fit the proposed model. Based on our experimental results, the best performance is achieved using a proposed model with a classification accuracy of 98.2% for early detection of common bacterial blight disease on a common bean leaf. The result of this study can help farmers and domain experts by detecting CBB disease at an early stage for the necessary treatments.

Plant phenotyping relevance

インゲン葉の画像から病徴（健全／斑点細菌病）を推定する深層学習モデルを開発・評価しており、植物病害状態の取得・抽出が研究の中心である。

Code and data availability

The paper describes a self-collected dataset of 3,135 common bean leaf images and a DenseNet-based CNN model, but no block contains any public deposit, availability statement, or authors' URL for the dataset, code, or trained model. All URLs present are cited references, not paper-specific assets.

No evidence-backed public reproduction asset is currently recorded.