CNN and vision transformer-based hybrid ensemble learning for sugarcane leaf disease classification

Sugarcane leaf diseases can adversely affect crop productivity, motivating the development of reliable image-based classification methods for automated disease recognition. A systematic study of four convolutional neural networks (CNNs) – EfficientNetB0, DenseNet121, ResNet50 and VGG16 with a modified classification head – and two Transformer architectures – ViT-B/16 and Swin-B/7 – is presented under a unified experimental setting. A dataset of 19,926 images across six classes, including five disease categories and a healthy-leaf category, was used in the study, and included duplicate and near-duplicate auditing, source-family grouping, stratified partitioning, architectural ablation, prediction-diversity analysis, calibration assessment, statistical comparison, and computational complexity evaluation. DenseNet121 performed best with a test accuracy of 94.13% and macro-F1 score of 93.58% among the individual models. A six-model probability-level weighted-fusion method obtained 95.00% accuracy and a macro-F1 score of 94.52%, and the CatBoost meta-learner achieved an accuracy of 95.10%. The accuracy difference between the weighted ensemble and the CNN-only ensemble and between the CatBoost and weighted fusion was not statistically significant, but the improvement of the weighted ensemble over DenseNet121 was statistically significant, using the paired bootstrap confidence interval and exact McNemar’s test with Holm correction. The prediction-diversity analysis revealed measurable differences between the base models; prediction diversity alone does not establish robustness to distribution shift. Further, the computational analysis showed that there was a tradeoff between predictive performance and computational requirements. The results show that CNN, Transformer, and Ensemble models perform well for the classification of sugarcane leaf diseases, while emphasizing the need for statistical reliability, calibration, prediction behaviour, and computational cost.

Authors

Institutions

Publication Details

Journal
Discover Computing
Published
2026-11-21
DOI
https://doi.org/10.1007/s10791-026-10805-7
Primary Topic
Smart Agriculture and AI
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

CNN and vision transformer-based hybrid ensemble learning for sugarcane leaf disease classification

Korhan Cengiz, H. Varun Chand, Weiwei Jiang, M. N Sardar Islam
Discover Computing
Smart Agriculture and AI
article

CNN and vision transformer-based hybrid ensemble learning for sugarcane leaf disease classification

Korhan Cengiz, H. Varun Chand, Weiwei Jiang, M. N Sardar Islam
article en

Abstract

Sugarcane leaf diseases can adversely affect crop productivity, motivating the development of reliable image-based classification methods for automated disease recognition. A systematic study of four convolutional neural networks (CNNs) – EfficientNetB0, DenseNet121, ResNet50 and VGG16 with a modified classification head – and two Transformer architectures – ViT-B/16 and Swin-B/7 – is presented under a unified experimental setting. A dataset of 19,926 images across six classes, including five disease categories and a healthy-leaf category, was used in the study, and included duplicate and near-duplicate auditing, source-family grouping, stratified partitioning, architectural ablation, prediction-diversity analysis, calibration assessment, statistical comparison, and computational complexity evaluation. DenseNet121 performed best with a test accuracy of 94.13% and macro-F1 score of 93.58% among the individual models. A six-model probability-level weighted-fusion method obtained 95.00% accuracy and a macro-F1 score of 94.52%, and the CatBoost meta-learner achieved an accuracy of 95.10%. The accuracy difference between the weighted ensemble and the CNN-only ensemble and between the CatBoost and weighted fusion was not statistically significant, but the improvement of the weighted ensemble over DenseNet121 was statistically significant, using the paired bootstrap confidence interval and exact McNemar’s test with Holm correction. The prediction-diversity analysis revealed measurable differences between the base models; prediction diversity alone does not establish robustness to distribution shift. Further, the computational analysis showed that there was a tradeoff between predictive performance and computational requirements. The results show that CNN, Transformer, and Ensemble models perform well for the classification of sugarcane leaf diseases, while emphasizing the need for statistical reliability, calibration, prediction behaviour, and computational cost.

Discover ComputingVol. 29(1)
Beijing University of Posts and Telecommunications (CN), APJ Abdul Kalam Technological University (IN), University of Khorfakkan (AE), Victoria University (AU)
Smart Agriculture and AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon