Custom Convolutional Neural Network for Detection and Classification of Rice Plant Diseases

Abstract

Agriculture and its allied sector provide the livelihoods of vast rural areas in India. It also plays a vital role in the growth of developing countries like India. Rice, one of the primary components of Indian agriculture plays a significant role in the food security of the country and the world. However, it is stricken by many diseases that affect the quality and quantity of rice yield. This paper proposes a custom CNN architecture for detecting and classifying common diseases found in rice plants by reducing the number of parameters associated with the network. The proposed CNN architecture has been trained using a dataset of four types of common rice plant diseases. In addition, 1400 on-field images of a healthy rice leaf image dataset have been introduced in the paper for the detection of disease-free plants as well. Independent experiments were carried out with and without the inclusion of the healthy leaf image dataset. The performance of the proposed model is evaluated with the use of Stochastic Gradient Descent with Momentum (SGDM) and Adaptive Moment Estimation (Adam) optimization techniques using several performance matrices. Experimental results from the dataset for the classification of four rice plant diseases show that the model using SGDM optimization gives maximum accuracy of 99.66% and the model using the Adam optimization gives maximum accuracy of 99.83% on the test set in the 7th epoch respectively. However, the model with Adam optimizer performed better than the model with SGDM optimizer when the healthy leaf image dataset is included giving a maximum accuracy of 99.66% and 97.61% in the 7th epoch respectively.