Resilience in Remote Sensing Image Classification: Evaluating Deep Learning Models Against Adversarial Attacks
P. Hemashree, G. Padmavathi · 2024
In the domain of Remote Sensing Image Classification, achieving high accuracy is of utmost importance. This proposed article embarks a wide-ranging comparative study, employing the state-of-the-art deep learning models including ResNet, DenseNet, EfficientNet, VGG16, and InceptionV3, precisely evaluated on the remote sensing datasets such as EuroSAT, UCMerced-LandUse, and NWPURESISC45. Traditional evaluation metrics, including Accuracy, Precision, F1 Score, Recall and Loss are used to evaluate the test data. The study reveals that DenseNet model achieves the highest accuracy across all datasets with an accuracy of $\mathbf{9 7. 8 8 \%}$ on EuroSAT, 95.95% on UCMerced-LandUse, and $89.80 \%$ on NWPU-RESISC45, demonstrating its robust performance without attacks. Extending the analysis, adversarial attacks such as Fast Gradient Sign Method (FGSM), Iterative - Fast Gradient Sign Method (I-FGSM), Projected Gradient Descent (PGD), and Carlini and Wagner (C&W) are introduced to assess the deep learning-based models’ robustness. The attacks on the aforementioned datasets reveal insights into the models’ vulnerabilities and resilience. The DenseNet model that performed well before attacks was compromised after the attacks with an accuracy of $54.14 \%$ on EuroSAT, $75.29 \%$ on UCMerced-LandUse, and $41.20 \%$ on NWPU-RESISC45, revealing its vulnerability to attacks. The other models also show a significant decrease in accuracy after attacks, highlighting the need for robustness enhancement in remote sensing image classification. This research forms the groundwork for establishing the fact that the introduction of adversarial attacks reduce the classification accuracy considerably and enhancing the classification accuracy in remote sensing images by fortifying model robustness is an important pursuit in a period where remote sensing technology expands its role in environmental monitoring, calamity management, urban planning and development, climate change studies and other applications demanding reliable and accurate classification systems.