OryzaNet: Leaf Quality Assessment of Oryza sativa Using Hybrid Machine Learning and Deep Neural Network

Plant diseases are the major cause of agricultural production losses. Visual leaf assessment based on human capability is the traditional diagnostic approach but leads to errors that contribute to losses. As a response, the analysis of rice (Oryza sativa) leaf quality through the development of a diagnostic method based on computational imaging was explored in this study. Feature-based machine learning (ML) algorithms including k-nearest neighbors (KNN), linear discrimination analysis (LDA), Naïve Bayes (NB), decision tree for classification (CT), and support vector machine (SVM) were configured for the classification of the health status of rice leaves according to the spectro-textural characteristics of image samples. A total of 17 features were extracted through the HSV color space thresholding technique which was further simplified through neighborhood-principal component analysis (NCA-PCA) selection that resulted in a four-feature vector (a*, Cr, contrast, homogeneity). Further, Deep neural networks (DNN) comprising ResNet-50, ResNet-101, GoogLeNet, MobileNetv2, and Inceptionv3 networks were used to categorize the disease variants. As a result, all ML models exhibited excellent classification performance. Additionally, the KNN model delivered an accuracy rating of 97.92% for health condition classification. Also, among the created DNN models, DNN
RN50
produced a disease identification accuracy of 95.83%. With that, a vision-based diagnostic tool, OryzaNet, was established. This technique is a non-intrusive and novel approach to phenotyping and assessing the quality of rice leaves.