Evaluating Machine Learning, Deep Learning, and YOLO Models with Preprocessing and Feature Extraction Variations for Melon Leaf Disease Detection

Abstract

In the idea of creating a digital village system, the digital transformation of the agricultural sector is crucial. A crucial step within the agricultural process is recognizing plant diseases. Automatic identification of plant diseases is essential to streamline agricultural procedures. In light of this, the study aims to assess different machine learning, deep learning, and YOLO (You Only Look Once) techniques for identifying melon diseases through the use of leaf images. In the experiments, data preparation involving preprocessing and feature extraction tasks was conducted using various methods, both manually and automatically through the use of pretrained models like ResNet50 (50-Layer Residual Network). According to the findings, various machine learning techniques utilizing features obtained from a pretrained ResNet50 model achieved optimal outcomes with an accuracy reached up to 100%. The preprocessing set and the extracted features assured the identification of the disease characteristics. Among the techniques, Gradient Boosting, Naïve Bayes, Random Forest, and Support Vector Machine are suggested. Modeling using machine learning with manually set features, Convolutional Neural Networks, ResNet50, and YOLO showed accuracy levels that remain relatively low, ranging from 54.5% to 81.9%. For all subsequent techniques, enhancing preprocessing, further extracting features, improving labeling and bounding box configuration accuracy, managing data imbalance, applying augmentation variations, and thorough parameter tuning could be advantageous.