Lightweight Convolutional Neural Network Model Based on Modified MobileNetV3 for Plant Disease Classification

Abstract

Plant diseases remain a serious obstacle to global food production, leading to yield losses and reduced agricultural efficiency. Although deep learning approaches have shown strong performance in plant disease classification, their computational demands often prevent use on low-power or mobile devices. To address this limitation, we propose a compact deep learning model derived from a modified MobileNetV3 framework. The design incorporates sandglass and inverted residual blocks, enabling effective feature extraction while keeping the parameter count minimal. Performance was assessed on four benchmark datasets, PlantVillage, Apple, Maize, and Rice, where the model obtained accuracy scores of 93.68%, 84.38%, 90.12%, and 96.66%, respectively, using only 0.4 million parameters. The key contributions of this work are: (1) development of a lightweight architecture specifically optimized for plant disease classification, (2) substantial reduction in model complexity compared to standard MobileNetV3 without loss of accuracy, and (3) validation across multiple crop datasets to demonstrate robustness. These outcomes indicate that the proposed model is highly suitable for real-world agricultural applications in computationally constrained environments.