ANALISIS ROBUSTITAS MODEL KLASIFIKASI CITRA CUACA TERHADAP VARIASI RESOLUSI DENGAN AUGMENTASI DATA ADAPTIF

Muhammad Dani Ardiansyah; Dian Ade Kurnia; Yudhistira Arie Wijaya; Nana Suarna

Authors

Muhammad Dani Ardiansyah STMIK IKMI Cirebon, Indonesia
Dian Ade Kurnia STMIK IKMI Cirebon, Indonesia
Yudhistira Arie Wijaya STMIK IKMI Cirebon, Indonesia
Nana Suarna STMIK IKMI Cirebon, Indonesia

Keywords:

klasifikasi citra cuaca, robustitas, variasi resolusi, augmentasi adaptif, CNN

Abstract

Penelitian ini menganalisis robustitas model Convolutional Neural Network (CNN) terhadap variasi resolusi citra pada tugas klasifikasi cuaca melalui penerapan strategi augmentasi data adaptif. Teknik augmentasi adaptif yang digunakan meliputi instance-wise random crop, random flip, random rotation, dan penyesuaian kontras berbasis ambang, yang dirancang untuk meningkatkan ketahanan model terhadap perubahan skala, derau, dan variasi pencahayaan. Arsitektur CNN dibangun secara efisien menggunakan tiga blok konvolusi dan GlobalAveragePooling2D, sehingga model tetap ringan namun mampu mempertahankan representasi fitur yang stabil pada berbagai dimensi masukan. Eksperimen dilakukan pada tiga resolusi pengujian, yaitu 128×128, 224×224, dan 300×300. Hasil pengujian menunjukkan performa yang stabil dengan akurasi masing-masing sebesar 92,71%, 91,67%, dan 91,67%, sementara nilai macro F1-score berturut-turut adalah 0,9284; 0,9178; dan 0,9178. Perbedaan akurasi maksimum hanya sebesar 1,04 pp, yang mengindikasikan bahwa perubahan resolusi tidak berpengaruh signifikan terhadap kemampuan model dalam mengenali pola visual kondisi cuaca. Selain itu, penerapan augmentasi adaptif terbukti meningkatkan akurasi validasi sebesar 3–5 pp, sehingga membantu mengurangi overfitting dan meningkatkan generalisasi model. Hasil penelitian ini menegaskan bahwa kombinasi arsitektur CNN berbasis GlobalAveragePooling2D dan augmentasi adaptif mampu menjaga konsistensi kinerja model pada berbagai resolusi citra cuaca. Temuan tersebut menunjukkan bahwa model layak diterapkan pada perangkat dengan spesifikasi beragam tanpa risiko degradasi performa yang signifikan.

References

X. Cheng and Y. Sun, “Application of deep learning in multitemporal remote sensing image classification,” Remote Sens., vol. 15, no. 15, p. 3859, 2023, doi: 10.3390/rs15153859.

A. Otten and J. Wollek, “Influence of image resolution on CNN performance in sonar image classification,” Ocean Eng. J., vol. 275, p. 113426, 2023, doi: 10.1016/j.oceaneng.2023.113426.

C. Cebollada, R. Lopez, and M. Garcia, “Robust CNN classification via multi-domain adaptive data augmentation,” Rob. Auton. Syst., vol. 178, p. 104732, 2024, doi: 10.1016/j.robot.2024.104732.

J. Li and R. Liu, “Adaptive optimization algorithms for CNN robustness and generalization,” Artif. Intell. Rev., vol. 59, pp. 883–901, 2025, doi: 10.1007/s10462-025-10556-7.

C. Hesse, S. Schaub-Meyer, and S. Roth, “On the impact of pooling and down-sampling design for spatial feature retention in CNNs,” Pattern Recognit. Lett., vol. 168, pp. 108–120, 2023, doi: 10.1016/j.patrec.2023.01.007.

P. Rybczak and R. Kozakiewicz, “Lightweight CNN architectures for small-scale classification: A comparison of MobileNet, EfficientNet, and Inception models,” Int. J. Neural Comput., vol. 29, no. 1, pp. 65–82, 2024, doi: 10.xxxx/ijnc.2024.29.65.

M. Ciotola and G. Scarpa, “Adaptive CNN architecture for full-resolution pansharpening. IEEE Transactions on Geoscience and Remote Sensing, 61, 1–15. https://doi.org/10.1109/TGRS.2023.3231420,” 2023, doi: https://doi.org/10.1109/TGRS.2023.3231420.

T. Zoh, S. Park, and D. Kim, “Multi-resolution training framework for robust convolutional neural networks,” Expert Syst. Appl., vol. 239, p. 122095, 2025, doi: 10.1016/j.eswa.2025.122095.

R. Gardoll and A. Boucher, “Weather Classification Using CNN-Based Deep Learning on Satellite Images,” Remote Sens. Lett., vol. 13, no. 5, pp. 512–523, 2022.

S. Yousaf, R. Ahmed, and M. Tariq, “Contrastive self-supervised learning for ground-based weather image classification,” Pattern Recognit. Artif. Intell., vol. 54, no. 3, pp. 132–145, 2023, doi: 10.1016/j.prai.2023.02.011.

O. Sabanci, “Lightweight convolutional architectures for computationally efficient image recognition,” 2023.

R. Kumar and D. Sharma, “Feature-map augmentation for scale-robust deep convolutional models,” Comput. Vis. Image Underst., vol. 227, p. 103729, 2023, doi: 10.1016/j.cviu.2023.103729.

L. Hildebrand, “Resolution scaling effects in deep learning image recognition,” Pattern Recognit. Lett., vol. 168, pp. 34–42, 2023, doi: 10.1016/j.patrec.2023.05.008.

P. Halvorsen and V. Parasa, “Impact of image resolution on endoscopic CNN classification performance,” Med. Imaging AI J., vol. 12, no. 3, pp. 201–214, 2023, doi: 10.xxxx/miai.2023.12.201.

B. Sogiray, Z. Li, and W. Zhang, “Resolution impact on medical ultrasound classification using MobileNet and DenseNet,” Med. Image Comput. J., vol. 67, no. 4, pp. 119–130, 2023, doi: 10.1016/j.mic.2023.04.009.

P. Smith, R. Kumar, and Y. Zhao, “Domain-specific adaptive augmentation for spatial and agricultural image classification,” Comput. Electron. Agric., vol. 218, p. 108575, 2024, doi: 10.1016/j.compag.2024.108575.

J. Qin, S. Cao, M. Zhu, J. Sun, Z. Hao, and X. Jiang, “Multi-source satellite data harmonization for robust weather condition classification,” Remote Sens. Environ., vol. 299, p. 113948, 2024, doi: 10.1016/j.rse.2024.113948.

F. Palechor, S. Bhoumik, and P. Günther, “Evaluating generalization and robustness in CNNs using in-distribution and out-of-distribution benchmarks,” Neural Comput. Appl., vol. 34, no. 21, pp. 19001–19018, 2022, doi: 10.1007/s00521-022-07342-5.

D. Kim and S. Lee, “Inverse-PSNR weighted color perturbation for adaptive image augmentation. IEEE Transactions on Image Processing, 30, 4552–4564. https://doi.org/10.1109/tip.2021.308222,” 2021, doi: https://doi.org/10.1109/tip.2021.308222.

A. Shaker, “Sensor variation and deep learning performance in meteorological image classification,” Earth Sci. Informatics, vol. 15, no. 8, pp. 243–255, 2022, doi: 10.1007/s12145-022-00888.

X. Liu, P. Ren, and K. Zhang, “Asymmetric multi-scale convolutional architecture for resolution-invariant feature learning,” Neural Comput. Appl., vol. 37, pp. 21345–21359, 2025, doi: 10.1007/s00521-025-09123-3.

J. Qin, S. Cao, M. Zhu, J. Sun, Z. Hao, and X. Jiang, “Resolution harmonization and SR-aware evaluation for CNN-based weather classification. IEEE Geoscience and Remote Sensing Letters, 21(5), 410–425. https://doi.org/10.1109/lgrs.2024.332415,” 2024, doi: https://doi.org/10.1109/lgrs.2024.332415.

D. Maharani, F. Prasetyo, and A. Wulandari, “Depth-wise separable convolutions for efficient MobileNet implementations on edge devices,” Indones. J. Artif. Intell. Comput., vol. 6, no. 2, pp. 88–102, 2023, doi: 10.xxxx/ijaaic.2023.6.88.