Progressive Resolution Training with Attention-Enhanced Ensemble Learning for Automated Recognition of Sugarcane Leaf Diseases
Diyala Journal of Engineering Sciences · 15.06.2026 · 15-34
DOI: 10.24237/djes.2026.19202 ↗Kurz gefasst
Fachartikel in der Zeitschrift Diyala Journal of Engineering Sciences, online erschienen am 15.06.2026. Verfasst von Gautam Kumar und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Sugarcane (Saccharum officinarum) is one of the most economically important crops in the world. However, outbreaks of leaf diseases cause extensive annual yield losses and threaten sugarcane agricultural productivity. Therefore, early and accurate disease detection is vital for effective crop management, yield protection, and food security. This study presents a deep ensemble learning framework, termed CBAM-SugarcaneNet, for the automated recognition of sugarcane leaf diseases under field conditions. The proposed framework integrates progressive multi-resolution training, a Convolutional Block Attention Module (CBAM), focal loss, Sharpness-Aware Minimization (SAM), ensemble learning, and test-time augmentation to address major challenges such as class imbalance, inter-class visual similarity, background variation, and limited training data. A three-stage progressive training strategy was employed using ConvNeXt-Large and EfficientNet-B5/B7 architectures enhanced with CBAM attention modules to improve the accuracy of disease-relevant feature extraction. Focal loss was used to emphasize minority and difficult-to-classify disease classes, whereas SAM improved model generalization by encouraging convergence toward flatter minima. Experiments were conducted on an 11-class sugarcane leaf disease dataset comprising 6,748 images. The proposed framework achieved 97.2% accuracy for single-model deployment, 97.6% for the ensemble configuration, and 98.1% with test-time augmentation, with a mean 5-fold cross-validation accuracy of 98.05% ± 0.14%. Grad-CAM visualizations further confirmed that the model focused on disease-affected leaf regions. The developed framework offers a robust and deployable solution for precision-agriculture-based disease diagnosis in sugarcane.
Zitieren
APA
Kumar, G., & Bhatnagar, V. (2026). Progressive Resolution Training with Attention-Enhanced Ensemble Learning for Automated Recognition of Sugarcane Leaf Diseases. Diyala Journal of Engineering Sciences, 15-34. https://doi.org/10.24237/djes.2026.19202
Vancouver
Kumar G, Bhatnagar V. Progressive Resolution Training with Attention-Enhanced Ensemble Learning for Automated Recognition of Sugarcane Leaf Diseases. Diyala Journal of Engineering Sciences. 2026:15-34. doi:10.24237/djes.2026.19202
BibTeX
@article{Kumar2026,
author = {Kumar, Gautam and Bhatnagar, Vivek},
title = {Progressive Resolution Training with Attention-Enhanced Ensemble Learning for Automated Recognition of Sugarcane Leaf Diseases},
journal = {Diyala Journal of Engineering Sciences},
year = {2026},
pages = {15--34},
doi = {10.24237/djes.2026.19202}
}
EUVIMED