Classification of pepper leaf diseases using VGG16.NET

Agriculture is one of the most essential resources for sustaining human life and fostering the development and productivity of nations. To meet increasing demands for vegetables and fruits, it is imperative to conduct research and implement preventive measures to enhance production efficiency. Among these measures, addressing plant diseases stands out as a top priority. Recognizing and diagnosing plant diseases is critical for improving agricultural productivity, preventing the spread of infections, and minimizing economic losses. This study aims to classify and diagnose pepper leaf diseases, enabling early intervention to mitigate their impact. Diseases such as mildew, mites, caterpillars, worms, aphids, and leaf burn predominantly affect greenhouse peppers, leaving distinctive marks on their leaves. These unique features form the basis for developing a 7-class classification model that facilitates faster and more accurate diagnosis of pepper leaf diseases. While traditional direct measurement methods are straightforward and reliable, they are often time-consuming and labor-intensive. To overcome these limitations, Convolutional Neural Network (CNN) algorithms for image processing have been employed. A key preprocessing step involves color enhancement to amplify variations in green and yellow hues, improving the differentiation between disease classes. An innovative algorithm has been designed to enhance the vibrancy, contrast, and overall color properties of the images, ensuring optimal feature extraction. The classification process leverages the 19-layer VGGNet architecture, renowned for its high accuracy in handling complex datasets. To reduce training time, the pre-trained layers of VGGNet were frozen, and additional layers were added for fine-tuning. The proposed model was evaluated on a proprietary dataset specifically compiled for this study. To the best of our knowledge, this is the first study to focus on diagnosing aphid and caterpillar classes within this context. The model achieved an average accuracy of 92.00%, considered highly satisfactory for a 7-class diagnostic task. The primary source of misclassification was observed in the aphid class, attributed to the limited number of samples available for training.