Ilusi Keterjelasan Kuantifikasi Penjajaran Grad-CAM: Studi Kasus Masker Segmentasi Warna Otomatis Citra Penyakit Padi

Agung Ramadhanu; Rely  Maryani

doi:10.35134/komtekinfo.v10i4.719

Authors

Agung Ramadhanu Universitas Putra Indonesia YPTK Padang
Rely Maryani Universitas Putra Indonesia YPTK Padang

DOI:

https://doi.org/10.35134/komtekinfo.v10i4.719

Keywords:

Explainable AI, Grad-CAM, Attention Alignment, Segmentation Mask, Rice Leaf Disease

Abstract

Saliency maps are widely used to explain convolutional networks in plant disease studies, yet they are usually shown as a few selected images without numbers, and the numbers that do exist depend on reference masks that are rarely validated. This diagnostic case study pairs a visual saliency map with quantitative alignment scores for one image per class and checks whether an automatic reference mask is fit for that purpose. Grad-CAM saliency was computed at the last convolutional block (7 × 7) of an ImageNet-pretrained EfficientNet-B0 that serves as a negative control: it was not fine-tuned on rice images and the target is the highest ImageNet logit. The first image of each class in a deduplicated public Kaggle dataset was used (four classes, four images). Reference masks were obtained by unsupervised CIELAB K-Means (k = 3) inside an HSV/Otsu leaf mask and are intended to mark lesions. Three measures were computed: Lesion Energy Hit Rate (H_E), binary Intersection-over-Union (IoU_B, threshold 0.4), and a Normalized Energy Gain (G), the difference between H_E and the chance level implied by mask area. H_E was 63.4% (Blast), 52.9% (Brown Spot), 45.2% (Bacterial Blight), and 1.9% (Healthy); IoU_B was 42.1%, 33.9%, 10.6%, and 2.9% for Brown Spot, Bacterial Blight, Blast, and Healthy, respectively. G was −13.4 percentage points for Blast and between +0.7 and +12.1 for the other three images (mask areas measured from the figure, hence approximate). Visual inspection showed that in these images the masks largely cover green leaf area and not the visible lesions (for example, the brown spot in the Blast image lies outside the mask). The scores therefore measure alignment with colour-cluster masks, cannot be read as lesion alignment, and illustrate an interpretability illusion when the mask is not validated. With one image per class, they illustrate cases and are not class averages

References

R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-CAM: Visual explanations from deep networks via gradient-based localization,” Int. J. Comput. Vis., vol. 128, no. 2, pp. 336–359, 2020, doi: 10.1007/s11263-019-01228-7.

A. Chattopadhay, A. Sarkar, P. Howlader, and V. N. Balasubramanian, “Grad-CAM++: Generalized gradient-based visual explanations for deep convolutional networks,” in Proc. IEEE Winter Conf. Appl. Comput. Vis. (WACV), 2018, pp. 839–847, doi: 10.1109/WACV.2018.00097.

M. M. Rahman et al., “Explainable deep transfer learning framework for rice leaf disease diagnosis and classification,” Int. J. Adv. Comput. Sci. Appl. (IJACSA), vol. 15, no. 12, pp. 862–884, 2024.

S. M. Alhammad et al., “Deep learning and explainable AI for classification of potato leaf diseases,” Front. Artif. Intell., vol. 7, Art. no. 1449329, 2025, doi: 10.3389/frai.2024.1449329.

P.-J. Kindermans et al., “The (un)reliability of saliency methods,” arXiv:1711.00867, 2017.

M. A. Noyan, “Uncovering bias in the PlantVillage dataset,” arXiv:2206.04374, 2022.

J. Zhang, S. A. Bargal, Z. Lin, J. Brandt, X. Shen, and S. Sclaroff, “Top-down neural attention by excitation backprop,” Int. J. Comput. Vis., vol. 126, no. 10, pp. 1084–1102, 2018, doi: 10.1007/s11263-017-1059-x.

L. Arras, A. Osman, and W. Samek, “CLEVR-XAI: A benchmark dataset for the ground truth evaluation of neural network explanations,” Inf. Fusion, vol. 81, pp. 14–40, 2022, doi: 10.1016/j.inffus.2021.11.008.

A. Hedström et al., “Quantus: An explainable AI toolkit for responsible evaluation of neural network explanations and beyond,” J. Mach. Learn. Res., vol. 24, no. 34, pp. 1–11, 2023.

J. Adebayo, J. Gilmer, M. Muelly, I. Goodfellow, M. Hardt, and B. Kim, “Sanity checks for saliency maps,” in Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 31, 2018, arXiv:1810.03292.

A. Saporta et al., “Benchmarking saliency methods for chest X-ray interpretation,” Nat. Mach. Intell., vol. 4, no. 10, pp. 867–878, 2022, doi: 10.1038/s42256-022-00536-x.

M. Tan and Q. Le, “EfficientNet: Rethinking model scaling for convolutional neural networks,” in Proc. 36th Int. Conf. Mach. Learn. (ICML), PMLR 97, 2019, pp. 6105–6114.

O. Russakovsky et al., “ImageNet large scale visual recognition challenge,” Int. J. Comput. Vis., vol. 115, no. 3, pp. 211–252, 2015, doi: 10.1007/s11263-015-0816-y.

“Rice Leaf Disease Detection Dataset,” Kaggle, pengguna shams069. [Online]. Available: https://www.kaggle.com/datasets/shams069/rice-leaf-disease-detection-dataset

R. Maryani and [Nama Dosen Pembimbing], “Fusi fitur deep dan fitur buatan tangan berbasis masker warna dengan seleksi F-ANOVA pada klasifikasi penyakit daun padi: evaluasi terkontrol kebocoran data,” [jurnal dan status: diisi penulis], 2026.