CottonTriFusionNet: Explainable Multibackbone Feature Fusion With SE Attention for Cotton Leaf Disease Classification
Applied Computational Intelligence and Soft Computing · 2026-01 · 2026 (1)
DOI: 10.1155/acis/4429407 ↗Kurz gefasst
Fachartikel in der Zeitschrift Applied Computational Intelligence and Soft Computing, online erschienen am 10.11.2026. Verfasst von Md. Saymon Hosen Polash und 4 weiteren Autor:innen. Die Arbeit nennt 37 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Woosong University.
Abstract (Originalsprache)
Cotton leaf diseases reduce crop yield, affect fiber quality, and create economic losses for farmers and the textile sector. This study proposes CottonTriFusionNet, a squeeze‐and‐excitation (SE) attention‐enhanced multibackbone framework that combines three pretrained convolutional neural network (CNN) backbones: EfficientNetB3, ResNet50, and DenseNet201. The SE‐refined 1536‐, 2048‐, and 1920‐dimensional features are concatenated into a 5504‐dimensional representation, from which a learnable attention module dynamically generates input‐dependent weights before multilayer classification. Two public cotton leaf disease datasets were used to evaluate the model. CottonTriFusionNet achieved an accuracy of 99.08% and a macro F1 score of 0.9913 on D1, and 99.17% accuracy and a 0.9917 macro F1 score on D2. Five‐fold cross‐validation additionally produced mean accuracies of 97.43% and 99.46% on D1 and D2, respectively. Reliability metrics, paired t ‐tests, and Wilcoxon signed‐rank tests with Holm correction, and robustness tests under noise, blur, occlusion, and CutMix provided additional evidence of the model’s reliability and sensitivity under challenging conditions. Ablation studies showed the contribution of SE attention, all three backbones, and the proposed fusion design. Cross‐dataset evaluation achieved 88.38% accuracy without target‐domain fine‐tuning. Gradient‐weighted class activation mapping (Grad‐CAM) visualization showed that the model focused mainly on disease‐affected areas of the leaves. A web‐based application was also implemented to demonstrate the practical deployment potential of the framework. Overall, CottonTriFusionNet provides an accurate and interpretable framework with promising reliability and cross‐dataset transfer potential.
Förderung
- Woosong University
Zitieren
APA
Polash, M. S. H., Banavath, V., Islam, M. N., Sumon, M. S. H., & Uddin, J. (2026). CottonTriFusionNet: Explainable Multibackbone Feature Fusion With SE Attention for Cotton Leaf Disease Classification. Applied Computational Intelligence and Soft Computing, 2026(1). https://doi.org/10.1155/acis/4429407
Vancouver
Polash MSH, Banavath V, Islam MN, Sumon MSH, Uddin J. CottonTriFusionNet: Explainable Multibackbone Feature Fusion With SE Attention for Cotton Leaf Disease Classification. Applied Computational Intelligence and Soft Computing. 2026;2026(1). doi:10.1155/acis/4429407
BibTeX
@article{Polash2026,
author = {Polash, Md. Saymon Hosen and Banavath, Venkateshwarlu and Islam, Md. Naimul and Sumon, Md. Shakhauat Hossan and Uddin, Jia},
title = {CottonTriFusionNet: Explainable Multibackbone Feature Fusion With SE Attention for Cotton Leaf Disease Classification},
journal = {Applied Computational Intelligence and Soft Computing},
year = {2026},
volume = {2026},
number = {1},
doi = {10.1155/acis/4429407}
}
EUVIMED