Adversarial Robustness of Deep Learning Models for Inland Water Body Segmentation from SAR Images

Segmentation of inland water bodies from Synthetic Aperture Radar (SAR) images is crucial for several applications, such as flood mapping and monitoring. While SAR sensors provide high-resolution, all-weather imagery, distinguishing water from water-like surfaces remains challenging due to the complex geometry of inland regions. Semantic segmentation models, including deep learning and transformer-based approaches, are increasingly used for land-water segmentation. However, manual annotation - commonly used to generate water masks as ground truth - is prone to label noise, particularly near land-water boundaries. Currently, there are no datasets and studies that capture such human annotation errors to help train models to be robust to such noise. In this work, we simulate manual annotation errors as adversarial attacks using morphological operations and study the robustness of various segmentation models on our curated dataset. We compare the robustness of the chosen models to these morphological poisoning attacks, and also create a publicly available dataset with these adversarial examples. Our results show that U-Net can tolerate a certain level of annotation corruption before a sharp performance drop, while other models exhibit a gradual decline, with Mask2Former being an anomaly due to the smaller dataset size and its mask-classification paradigm. We validate the statistical significance and robustness of the results through paired t-tests and 95% confidence intervals. We also discuss a few anomalies in the dataset itself and how the U-Net model is immune to these anomalies. These findings highlight the critical role of annotation quality in determining model effectiveness. We also release our code and a new dataset containing these adversarial examples for robust model training. (GitHub: https://github.com/GVCL/IWSeg-SAR-Poison.git).