SAR2HEIGHT: Height Estimation from A Single SAR Image via Layover Backscattering Reconstruction and Deep Learning

Height estimation from a single Synthetic Aperture Radar (SAR) image has demonstrated a great potential in real-time environmental monitoring and scene understanding. However, recovering 3D information from 2D image is a mathematics ill-posed problem. Moreover, in mountainous regions, severe layover causes signal aliasing and great loss of geometric information. This research presents a single-image SAR height estimation framework that explicitly addresses layover-induced distortions by integrating physics-based modeling with deep learning. The proposed approach first reconstructs SAR backscattering in layover regions by establishing a one-to-many mapping between radar slant-range pixels and ground cells using SAR imaging simulation and a coarse digital elevation model. Mixed backscattered energy in layover pixels is then reallocated to individual ground locations according to physically derived contribution ratios, yielding a reconstructed SAR image with a more rational radiometric distribution.

Based on the reconstructed SAR data, an enhanced U-Net architecture with attention and selective-kernel mechanisms is employed for height estimation. Large-kernel selective modules enable adaptive multi-scale feature extraction to capture both local terrain details and long-range topographic context, while efficient channel attention emphasizes height-relevant feature channels. In addition, sparse elevation priors and Euclidean distance maps are incorporated to further constrain the inversion process. The datasets are constructed using Sentinel-1A SAR imagery and ground truth height maps derived from the Shuttle Radar Topography Mission (SRTM). The study focuses on three distinct regions characterized with different topography: Yumen in Gansu Province, China; Shule Nanshan in Qinghai Province, China; and the San Juan National Forest in the United States.

Experiments conducted demonstrate that the proposed framework substantially improves height estimation accuracy compared with conventional single-image SAR methods. Specifically, the reconstruction module mitigates signal aliasing by establishing a one-to-many mapping between slant-range and ground cells, successfully restoring a rational backscattering distribution in layover areas. This restoration alone reduces RMSE of the estimated height by 5.6%, 24.1%, and 25.3% across the three datasets. Complementing this, the ASK-UNet leverages LSK and ECA modules to capture multi-scale features, further refining the estimation accuracy. Compared with the baseline network, the ASK-UNet yields additional RMSE reductions of 7.3%, 5.5%, and 8.3% respectively. Overall, experimental results demonstrate that mSAR2Height achieves state-of-the-art performance with a total RMSE reductions of 12.4%, 28.2%, and 31.4%. The results indicate that combining physics-based layover reconstruction with attention-guided deep learning provides an effective and reliable solution for single SAR image height estimation in complex terrain, with high potential for rapid mapping and disaster response applications.