Knowledge Transfer-based Network from Medium and High-Resolution SAR Imagery for Built-Up Extraction with Class-Imbalanced Data

With the rapid urbanization, building detection in synthetic aperture radar (SAR) images has become an increasingly important research topic. However, Built-Up Extraction faces two challenges: class imbalance and inconsistent resolution. In this paper, to address the challenges of class imbalance and varying resolutions, we adopt SegFormer as the main framework and introduce Lovász loss to handle class imbalance while using knowledge distillation to transfer knowledge between different resolutions. Experiments show our method can improve built-up extraction accuracy. Our model achieves 0.8424 IoU on the competition test set.