Magnitude-Based Weight-Pruned Automated Convolutional Neural Network to Detect and Classify the Plant Disease

Abstract

Agriculture systems are constantly vulnerable to pathogenic viruses and the diseases caused by them, posing a threat to a country’s food security. Farmers often find it challenging to find the diseases in the plants at an early stage before it destroys the plant completely. In the proposed research work, an intelligent deep convolutional neural network for leaf image classification is developed, which can recognize 38 different types of plant diseases that are prevalent in 14 unique plant species. According to the complexity of the classification problem, various hyperparameters such as the number of epochs, batch size, hidden layers for feature extraction, dropout layers for regularization, and the number of neurons in each dense layer have been carefully designed in such a way that the model is neither overfitting nor underfitting, thus building an optimized deep CNN model. The developed CNN model for plant disease detection has an overall accuracy of 95% on the validation dataset. Further, magnitude-based weight pruning is carried out to reduce the network size by 66.7% and the overall accuracy is increased by 2%. Out of 33 test images, the model has predicted the plant diseases with an overall accuracy of 93.9% on the previously unseen test dataset. Thus, farmers would be highly benefitted from the proposed less complex weight-pruned CNN model as it predicts plant diseases using the concept of feature extraction with high accuracy, if a diseased leaf image of a plant is given as an input.