What question did this study set out to answer?

Investigate optimal CNN and Vision Transformer fusion strategies and benchmark quantitative explainable AI methods for automated cotton leaf disease classification.

August 19, 2026ElectronicsOpen Access

Cotton Leaf Disease Detection via Dual-Backbone CNN-Transformer Fusion with Quantitative XAI Comparison

View Full Paper

Ask AI

Bookmark

Authors

NUNaeem Ullah IFIvanoe De Falco GSGiovanna Sannino

Discussion

Member takes

Overview

Computational evaluation demonstrates that CNN-Transformer feature fusion achieves 99.42% accuracy in cotton leaf disease classification, indicating effective integration for agricultural diagnostics.

Key Points

Investigate optimal CNN and Vision Transformer fusion strategies and benchmark quantitative explainable AI methods for automated cotton leaf disease classification.
Evaluated six individual backbone architectures (ResNet50, EfficientNet-B0, DenseNet121, MobileNetV2, ViT-Base, DeiT-Small) on a dataset of 1,711 cotton leaf images across four classes.
Systematically tested five dual-backbone fusion strategies (concatenation, attention, weighted, ensemble, and variance-based) alongside computational efficiency metrics.
Quantitatively benchmarked six explainable AI techniques (GradCAM, GradCAM++, ScoreCAM, LayerCAM, EigenCAM, and AblationCAM) using pointing game, IoU, AUC, and localization accuracy.
Individual baseline models attained highest accuracies of 92.40% for DenseNet121 and 96.49% for ViT-Base.
Concatenation fusion achieved peak performance with 99.42% accuracy (95% CI: 98.2–100%), a weighted F1-score of 0.994, and a Matthews correlation coefficient of 0.992.
EigenCAM yielded the highest overall quantitative explainability score among all evaluated attribution methods.

Cite This Study

Ullah et al. (2026) studied this question.

synapsesocial.com/papers/6a8563c403308d306e2d72c4 https://doi.org/10.3390/electronics15163650

View Full Paper

Ask AI

Bookmark

August 19, 2026ElectronicsOpen Access

Cotton Leaf Disease Detection via Dual-Backbone CNN-Transformer Fusion with Quantitative XAI Comparison

View Full Paper

Ask AI

Bookmark

Authors

NUNaeem Ullah IFIvanoe De Falco GSGiovanna Sannino

Discussion

Member takes

Overview

Key Points

Investigate optimal CNN and Vision Transformer fusion strategies and benchmark quantitative explainable AI methods for automated cotton leaf disease classification.
Evaluated six individual backbone architectures (ResNet50, EfficientNet-B0, DenseNet121, MobileNetV2, ViT-Base, DeiT-Small) on a dataset of 1,711 cotton leaf images across four classes.
Systematically tested five dual-backbone fusion strategies (concatenation, attention, weighted, ensemble, and variance-based) alongside computational efficiency metrics.
Quantitatively benchmarked six explainable AI techniques (GradCAM, GradCAM++, ScoreCAM, LayerCAM, EigenCAM, and AblationCAM) using pointing game, IoU, AUC, and localization accuracy.
Individual baseline models attained highest accuracies of 92.40% for DenseNet121 and 96.49% for ViT-Base.
Concatenation fusion achieved peak performance with 99.42% accuracy (95% CI: 98.2–100%), a weighted F1-score of 0.994, and a Matthews correlation coefficient of 0.992.
EigenCAM yielded the highest overall quantitative explainability score among all evaluated attribution methods.

Cite This Study

Ullah et al. (2026) studied this question.

synapsesocial.com/papers/6a8563c403308d306e2d72c4 https://doi.org/10.3390/electronics15163650

View Full Paper

Ask AI

Bookmark