Comparative Analysis of MobileNetV3-Large and Small for Corn Leaf Disease Classification

Abstract

Corn leaf disease represents a significant threat to agricultural productivity, capable of causing substantial economic losses in Indonesia. Conventional identification methods, which rely on visual observation by farmers, are frequently subjective, time-consuming, and inaccurate. This study conducts a systematic comparative analysis of two efficient Convolutional Neural Network (CNN) architecture variants, MobileNetV3-Large and MobileNetV3-Small, for the classification of four corn leaf conditions: Gray Leaf Spot, Common Rust, Northern Leaf Blight, and Healthy. The research further evaluates the influence of two prevalent optimizers, Adam and Stochastic Gradient Descent (SGD), to ascertain the most optimal model configuration through hyperparameter tuning. The models were trained and evaluated using a local image dataset from Sampang, Indonesia, comprising 4000 images. The methodology included image preprocessing, data augmentation, and hyperparameter tuning of the learning rate and batch size. The results demonstrate that both architectures achieved exceptionally high accuracy. The principal finding reveals that MobileNetV3-Small unexpectedly outperformed its larger variant, attaining a peak accuracy of 99.5% with the SGD optimizer, a learning rate of 0.01, and a batch size of 32. In comparison, MobileNetV3-Large reached a maximum accuracy of 99.0% under a similar configuration. These findings underscore the considerable potential of lightweight architectures for the development of rapid, accurate, and field-deployable plant disease diagnostic applications on mobile devices using deep learning.