Development of a transfer learning based modified VGGNet model with exponential linear unit activation function for maize disease classification

Abstract

Maize plant is globally grown and used as food for end consumers, and as raw material in industries for further processing. This plant is often faced with variety of diseases that affect the farm output. Different machine learning models were employed to automate the classification of these diseases where recently, Deep Learning (DL) models have shown a superior performance over other machine learning approach. But DL traditional activation function Rectified Linear Unit (ReLU) is faced with a dying neurons problem. This research curates maize dataset with four classes and modified Visual Geometry Group (VGG-19) and replaced the fifth convolutional layer with inception modules and substituted the ReLU function with Exponential Linear unit (ELU) function.  The modified VGG-19 had a training accuracy of 95.52% on the curated dataset and 92.50% on a publicly available plant village dataset. The modified model had a higher performance on the curated dataset due the superior quality of images of the dataset. The modified model also outperformed the VGG-19 for two reasons, the modified architecture of the DL, and the ELU activation used which mitigate the dying neurons faced by the ReLU. This research reveals the ELU robustness in learning dynamic nature of plant disease dataset over the ReLU. Also, the curated datasets serve as pivotal resource for evaluating algorithm and model performance in maize disease detection and classification tasks, thereby advancing research in agricultural technology and fostering improvements in crop management practices. This piece of work when incorporated into Personal Digital Assistant (PDA) devices will help farmers scan through their farms for an easy identification and classification of maize diseases with higher accuracy.