Performance and Interpretability Analysis of EfficientNetB3 for Guava Leaf Disease Classification Using Score-CAM
Devotion : Journal of Research and Community Service · 25.03.2024 · 5 (3) · 411-428
DOI: 10.59188/devotion.v5i3.25788 ↗Kurz gefasst
Fachartikel in der Zeitschrift Devotion : Journal of Research and Community Service, online erschienen am 25.03.2024. Verfasst von Tia Siti Hawa. Der Volltext ist frei zugänglich (Open Access, CC BY-SA 4.0).
Abstract (Originalsprache)
Background
Guava leaf image classification requires both performance evaluation and examination of the visual basis of model decisions.
Objective
This study analyses the performance of EfficientNetB3 and interprets its predictions using Score-Weighted Class Activation Mapping (Score-CAM).
Method
The GuavaLeafVision dataset covers seven classes of disease, pest damage, and healthy conditions. An audit of 1,050 images identified 189 pixel-duplicate copies, leaving 861 unique images, which were split per class into 600 training, 128 validation, and 133 test images in a single random split with one seed (42); near-duplicates were not controlled. ImageNet-pretrained EfficientNetB3 was trained through feature extraction and fine-tuning, and the model was selected by the lowest validation loss.
Results
Fine-tuning reduced the validation loss from 0.229648 to 0.110465. On the test set, the model achieved 97.74% accuracy (130/133; 95% Wilson confidence interval 93.58–99.23%), 97.75% macro-precision, 97.20% macro-recall, and 97.37% macro-F1, with three errors in 133 predictions; the lowest recall was 86.67% for Mealybug Pests. Score-CAM on nine examples showed activation on the leaf together with background or corner responses in some cases.
Conclusion
The results support internal classification and decision inspection but do not establish symptom localization or field generalization; repeated splits, near-duplicate control, and external testing are still required.
Zitieren
APA
Hawa, T. S. (2024). Performance and Interpretability Analysis of EfficientNetB3 for Guava Leaf Disease Classification Using Score-CAM. Devotion : Journal of Research and Community Service, 5(3), 411-428. https://doi.org/10.59188/devotion.v5i3.25788
Vancouver
Hawa TS. Performance and Interpretability Analysis of EfficientNetB3 for Guava Leaf Disease Classification Using Score-CAM. Devotion : Journal of Research and Community Service. 2024;5(3):411-428. doi:10.59188/devotion.v5i3.25788
BibTeX
@article{Hawa2024,
author = {Hawa, Tia Siti},
title = {Performance and Interpretability Analysis of EfficientNetB3 for Guava Leaf Disease Classification Using Score-CAM},
journal = {Devotion : Journal of Research and Community Service},
year = {2024},
volume = {5},
number = {3},
pages = {411--428},
doi = {10.59188/devotion.v5i3.25788}
}
EUVIMED