Support Vector Machine and Artificial Neural Network Models for Classification of Wheat and Mustard Genotypes
Loading...
Date
2021-10
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
CCSHAU, Hisar
Abstract
The aim of this study was to classify the wheat and mustard genotypes using discriminant analyses, artificial neural networks and support vector machine models. The secondary data of 302 wheat and 870 mustard genotypes for 14 morphological variables were used. The class variable grain yield was categorized into 3 classes in wheat dataset, which makes it a multiclass problem. While, mustard genotypes were categorized into binary classes on the basis of grain yield, oil content and combined variables.
In discriminant analyses, the performance of regularized discriminant analysis was higher than that of linear and quadratic discriminant analyses for both the datasets. Out of the three artificial neural network (ANN) models used for wheat dataset, training accuracy of resilient propagation was higher whereas less satisfactory results were obtained for radial basis function (RBF) network as compared to multi-layer perceptron (MLP) networks. But in mustard dataset, the training accuracies were notably high and testing accuracies were at par for RBF neural networks as compared to MLP networks. Out of the six kernels used for support vector machine (SVM) classification, RBF kernel outperformed all other kernel functions for both the datasets. Then the outputs of SVM paradigm with six kernels were combined in an Ensemble with Weighted Accuracy (EWA) model. The ensemble model provided high prediction accuracies for both the datasets in comparison to individual kernel classifiers.
The particle swarm optimization (PSO) technique has set more suitable parameters, provided higher classification accuracy in both the datasets. The ensemble model outperformed the others with 95.1% training accuracy followed by resilient propagation neural networks (94.7%) and PSO optimized SVM (94.2%) for wheat genotypes. While for testing data set, the EWA model and PSO optimized SVM performed well with 94.9% accuracy. The classification of mustard genotypes was found better with the grain yield as class variable followed by oil content. The ensemble model outperformed the other classifiers with 93.5% training accuracy followed by PSO optimized SVM (92.6%) and RBF neural networks (91.9%) for mustard genotypes. Whereas for testing dataset, highest accuracy of 92.6% was achieved with PSO optimized SVM followed by all neural network models (90.7%). The lowest accuracies were obtained with linear discriminant analysis for both the datasets.