Effects of SAR Resolution in Automatic Building Segmentation Using CNN

Successful deep learning (DL) models for Synthetic Aperture Radar (SAR) images mostly require high-resolution images. The lack of datasets for historical and publicly available SAR images in high-resolution quality limits DL applications for SAR. We simulate SAR resolution scaling by applying filters with varying strengths in the preprocessing step of the dataset. A segmentation model is trained on each dataset along with different model architecture and backbone. The Feature Pyramid Network (FPN) model achieved the best results while being robust for datasets with weak filtering.