Integration of modified pre-trained VGGNet with inception modules and vanishing gradient problem solution-based activation functions for rice disease classification

Abstract

Rice is one of the crops grown and consumed in almost every nation. However, the crop is challenged by variety of diseases which results in degraded and low quality of the end produce. This calls for automation of disease identification and classification. Recently, Deep Learning (DL) models got much attention due to its remarkable performance in classification of plants disease. However, three factors influence the performance of DL models. These factors include the nature of dataset used in training, the architecture of the model, and the type of activation function used in the model. This work digs into these three factors; first, a new high quality dataset for rice disease was collected over the course of this research work. Secondly, VGG-16 pre-trained model was modified to optimise the performance of the model on the new dataset for training. Inception modules were then added and Fully Connected (FC) layer was replaced with an average pooling layer. The Rectified Linear Unit (ReLU), which is the traditional activation function of DL models, faces the dying neurons problem. Different activation functions that solve the problem of dying neurons were applied to the modified architecture to determine which function performs best on the rice dataset. The model utilizing the Exponential Linear Unit (ELU) activation function achieved remarkable accuracies, with a training accuracy of about 92.0%. Such performance holds significant promise for early disease detection in agricultural settings, potentially leading to increased farm productivity.