Deep Learning for Radargrammetric DSM Generation: A StereoSAR Dataset and Multi-Scale Fusion Network

Kurzfassung

Radargrammetry is an important technique for digital surface model (DSM) reconstruction, but accurate disparity estimation from synthetic aperture radar (SAR) stereo images remains challenging due to speckle noise and geometric distortions. Despite the success of deep learning in disparity estimation for stereo matching, its application in stereoscopic SAR (StereoSAR) is still limited due to the lack of high-quality training data and task-specific models. To address this issue, this study develops a deep learning framework for radargrammetric DSM generation, integrating dataset construction and a multi-scale SAR stereo matching network. The StereoSAR4DSM dataset is developed using TerraSAR-X imagery and high-resolution aerial DSMs, with enhanced epipolar rectification and three SAR-driven augmentation strategies: multi-looking variation, random pixel sampling, and random elevation perturbation. These strategies enrich data diversity and support robust deep learning model training. Based on this dataset, we design the Multi-Scale StereoSAR Fusion Network (MSSFNet) that constructs pyramid cost volumes and progressively integrates multi-scale cost information. An attention-guided fusion mechanism and a disparity refinement module further enhance matching accuracy and restore fine terrain structures. Experimental results on two test areas with different imaging modes demonstrate that the deep learning model trained with the StereoSAR4DSM dataset outperforms traditional approaches, and the proposed MSSFNet achieves the highest accuracy compared with other deep learning methods. In addition, comparative experiments show that the SAR-driven enhancement strategies significantly improve data diversity and lead to more accurate disparity estimation. Overall, these findings confirm the effectiveness of the proposed framework and highlight the potential of deep learning-based StereoSAR methods for efficient and accurate DSM reconstruction.