Detection of bacterial leaf blight disease of cotton (Gossypium Hirsutum L.) Using convolution neural network (CNN): Simulations with support vector machine (SVM) and naïve bayes algorithms

Abstract

The plant diseases are becoming the major threats to the food security, because they cause severe damage to the crop and lead to yield loss. Therefore, identification of disease is pre-requisite for the implementation of proper, efficient and timely remedial measures to avoid yield losses. The present study involved the application of a hybrid approach integrating convolutional neural networks (CNNs) with two different algorithms viz. support vector machine (SVM) and the Naïve Bayes algorithms to detect the bacterial leaf blight disease of cotton (Gossypium hirsutum L.). The proposed algorithms were based on mini-batch size, max epochs and bias learning rate and were trained using 180 and 160 leaf images, respectively including infected and non-infected (healthy) plants to identify bacterial leaf blight disease of cotton using threefold image classification and five-fold cross validation strategy with a range of iterations (1-250). The proposed SVM algorithm achieved 91.1±0.5% accuracy, compared with 95.5±0.3% accuracy with Naïve Bayes algorithm. The mean precision value of 78.4±0.7% and 80.5±0.7% were achieved with SVM and Naïve Bayes algorithms, respectively. The accuracy and precision value of bacterial leaf blight disease detection were significantly and linearly related to each other (R²=0.918**; p<0.01 for SVM and R² = 0.904**; p<0.01 for Naïve Bayes algorithm). The mini-batch losses were considerably reduced to 0.3980 and 0.3365, respectively with SVM and Naïve Bayes algorithms at 250 iterations, compared with 1.8642 and 1.8201, respectively at 1 epoch and single iteration. The mini-batch accuracy was highest (93.24 and 93.98%, for SVM and Naïve Bayes algorithms, respectively) at 25 epoch and 250 iterations. The study highlights that the identification of bacterial leaf blight disease of cotton crop can accurately by detected proposed models; although the level of accuracy was higher for Naïve Bayes algorithm.