Optimization of MobileNetV2 Architecture Model Using Convolutional Neural Network Algorithm for Sugarcane Leaf Disease Classification

Abstract

Sugarcane leaf diseases pose a serious threat to agricultural productivity, directly impacting food security and economic stability worldwide. Although deep learning has been widely applied to plant disease classification, lightweight models such as MobileNetV2 often struggle to achieve high accuracy. This study aims to optimize the MobileNetV2 model to classify sugarcane leaf diseases more accurately and efficiently. The dataset used consists of 4,800 images categorized into six classes: Bacterial Blight, Healthy, Mosaic, Red Rot, Rust, and Yellow. Unlike transfer learning, which relies on pre-trained MobileNetV2 weights, this study manually redesigns the model architecture to improve feature-extraction efficiency and overall performance. The optimization process includes fine-tuning techniques, dropout regularization, and adaptive learning rate adjustments to improve classification accuracy and inference speed. Experimental results indicate that the optimized model achieves an accuracy of 98.5%, representing a significant improvement over the transfer learning approach. The restructuring of MobileNetV2 layers has been proven to enhance the model’s ability to learn discriminative features more effectively. Moreover, the optimized model is computationally lightweight, making it suitable for real-time deployment on mobile-based systems without compromising accuracy. In the future, research can focus on improving the model’s generalization by utilizing a larger dataset with a more diverse range of disease categories. Additionally, performance comparisons with other model architectures can be conducted to identify solutions that are not only more accurate but also achieve faster, more efficient training times.

Plant phenotyping relevance

サトウキビ葉の病徴を画像から分類する深層学習モデルの再設計・最適化が研究の中心であり、植物の病害状態を推定するフェノタイピング手法に該当する。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.