A Hybrid Approach to Rice Leaf Disease Detection Using Transfer Learning and SVM

Rice is a prominent crop in India, with the country ranking second globally in rice production. However, rice cultivation is severely affected by several diseases, impacting crop yield and quality. Early detection of these diseases can significantly mitigate their adverse effects. This study focuses on detecting primary rice leaf diseases: brown spot, tungro, bacterial blight, and blast. We explore various models, including recent publications, and develop hybrid approaches employing deep features from VGG16, VGG19, and ResNet50 models, alongside handcrafted features such as Local Binary Patterns (LBP) and Histogram of Oriented Gradients (HOG). These features combined with shallow machine learning classifiers like Random Forest and Support Vector Machine (SVM), are evaluated for rice leaf disease classification. Different combinations of these extracted features, with and without Principal Component Analysis (PCA), are tested using various classifiers to assess their impact on classification accuracy. The models that stack deep features using ResNet50 and handcrafted features using LBP, followed by dimensional reduction via PCA and transformation using SVM, achieved near-perfect metrics across F1 score, precision, accuracy, recall, and AUC. These results indicate strong generalization and balanced performance, outperforming methods published in recent literature.