Congratulations to our ESR 14 from MENELAOSᴺᵀ, Omid Ghozatlou, who successfully defended his PhD thesis on the topic of “Learning with Adversarial Samples for Earth Observation Multispectral Images ” on January 10, 2024 at the University POLITEHNICA of Bucharest under the supervision of Prof. Dr. Ing. Daniela COLTUC.
Thanks to the impressive computational capabilities of deep learning (DL), it offers numerous advantages and delivers outstanding results in the realm of earth observation (EO) image processing. Given the intricate nature and high dimensionality of remote sensing (RS) images, attempting to analyze this data without the aid of DL becomes an
insurmountable challenge. Nonetheless, leveraging DL for EO images comes with its fair share of challenges. Additionally, RS images are susceptible to various issues, including temporal variations, disparities in spatial and spectral resolution, data imbalances, and the presence of adversarial or other artifacts. Adversarial samples are a significant challenge in EO due to their capacity to deceive DL models and disrupt their accuracy. These samples can originate from natural occurrences like clouds, shadows, or image artifacts. Finding robust solutions for these challenges remains a top priority in EO, given the critical importance of accurate data analysis.
This study proposes novel solutions for mitigating the adversarial sample issue in RS image classification. Four strategies are introduced: Active Learning, Query-by- Example, Physics-aware deep models, and Synthetic data via Generative Adversarial Networks (GANs). Active Learning strategically selects informative samples to enhance the model’s performance. This approach allows the model to learn from data that it is uncertain about, thus improving decision-making. Query-by-Example aims to find the most similar image to a given query and optimize the network’s weights to separate adversarial samples from the query image in the latent space. This approach results in the model focusing on normal samples while identifying adversarial ones as outliers. Another noteworthy approach gaining popularity in RS research is Physics- aware deep models. These models incorporate the physical properties of EO data, guiding deep neural networks to better cluster and understand the data. By leveraging physical knowledge during the learning process, the model becomes more trustworthy and resilient against adversarial attacks. The final approach involves the use of GANs for generating synthetic satellite images to make the classifier robust against adversarial. GANs address the adversarial issue by training two sub-models in an adversarial learning
framework. To achieve this, a well-established and efficient StyleGAN is employed for Sentinel-1 Ocean SAR images, even with limited training data. We assess the classifier’s performance using GAN-generated data samples as the training set.
Recent Comments