Onion Diseases Classifications by using Deep Learning Techniques

Abstract

The majority of Ethiopians continue to rely heavily on the agricultural sector, which remains one of the most significant industries in the country. This sector faces numerous challenges, including plant diseases that negatively impact yield quality and productivity. Early detection and accurate diagnosis of onion health are crucial for preventing the spread of various plant diseases, ideally through the use of technology rather than manual labor. Onion crops play a vital role in Ethiopian agriculture, significantly contributing to food security and livelihoods. However, onion diseases have resulted in substantial losses and reduced yields. Traditional observation methods used by farmers and agricultural experts can be time-consuming, costly, and sometimes inaccurate. Deep learning techniques provide some of the most effective models for identifying plant diseases. Convolutional Neural Networks (CNNs) are a prominent approach, consisting of multiple processing layers that learn to represent images at varying levels of abstraction. These models have greatly advanced visual object recognition and image classification, making them well-suited for detecting and classifying onion leaf diseases. The primary objective of this study was to design and develop a deep learning-based model for identifying and classifying the three most critical onion leaf diseases: Downy Mildew, Onion Leaf Blight, and Bacterial Soft Rot. A total of 12,438 images, including augmented images across four categories, were collected and prepared for experimentation. This study specifically focuses on diseased and healthy onion leaves sourced from various agricultural research centers in Ethiopia. Experimental results demonstrate that the proposed ResNet50 model can effectively detect and classify the four classes of onion leaf diseases, achieving an impressive classification accuracy of 98.75%. This performance surpasses that of other classical deep learning models, including VGG19, DenseNet201, InceptionV3, and standard CNNs.