Sugarcane Plant Disease Classification Based on Leaf Image Using ConvNeXt V2 Deep Learning Model

Abstract

Sugarcane plant diseases pose a significant threat to agricultural productivity, yet early and accurate identification remains challenging for farmers due to the limitations of manual inspection. This study proposes a sugarcane leaf disease classification system using ConvNeXt V2 Tiny, a modern convolutional architecture with a Global Response Normalization (GRN) mechanism, combined with an ensemble Stratified K-Fold Cross Validation strategy (K=6) to improve generalization on real-world field data. A dataset of 2,948 leaf images spanning five classes (Red Rot, Mosaic, Rust, Yellow Leaf, and Healthy) was used, with field-collected images held out as a fixed test set. The ensemble model achieved a mean validation accuracy of 98.49% ± 0.58% across six folds and a test accuracy of 98.39% on 427 unseen field images, with macro-average precision, recall, and F1-score each reaching 98%. ConvNeXt V2 Tiny substantially outperformed ResNet-50 (87.35%) and EfficientNetV2-S (83.37%) under identical experimental settings, demonstrating superior generalization across the domain gap between curated and field data. The primary contribution of this study is the first application of ConvNeXt V2 Tiny with ensemble K-Fold strategy for sugarcane disease classification, offering high accuracy with moderate computational complexity (28.6M parameters) and practical deployability, as demonstrated through the SugarScan web application.

Plant phenotyping relevance

サトウキビ葉画像から病害状態を推定する画像ベースの表現型解析手法が研究の中心であり、モデル性能の検証・比較も実施しているため含める。

Code and data availability

The paper's phenotyping inputs include a public Kaggle dataset (Sugarcane Leaf Disease Dataset, SLD) of sugarcane leaf disease images used for training/validation, plus field-collected images. Only the Kaggle dataset qualifies as a paper-specific public asset with an authors' URL; no author analysis code, trained model