Enhanced Plant Disease Classification through Transfer Learning with Modified VGGNet and Inception Modules

Abstract

Advances in technology and high computational power have driven extensive use of Machine Learning (ML) and Deep Learning (DL) models in diverse fields, such as object recognition, intelligent transportation, and precision agriculture, with particular focus on plant disease classification. Literature highlights that DL models, initially designed for large datasets like ImageNet, often lack efficiency for tasks with smaller datasets and fewer classes, necessitating architectural modifications for better performance. This study addresses this issue by exploring optimized activation functions to overcome the vanishing gradient problem, which can impede model training and accuracy in small-scale datasets.  Specifically, this research modifies VGG-16 and VGG-19 architectures, replacing the fifth convolutional layer with an additional 3x3x512 convolutional layer and integrating the ELU activation function instead of Swish. Two inception layer modules and a global pooling layer were also added, followed by softmax for classification. Trained on curated maize and rice datasets, the proposed model achieved 98.40% accuracy, surpassing the Swish function’s 97.97%. This improvement underscores the ELU function's effectiveness in improving model performance, with practical applications in early plant disease detection, aiding precision agriculture by enabling timely intervention and yield protection.