Skip to main content

Penetration Bias Correction in X-band InSAR DEMs (Greenland)

A hybrid AI-physical method to correct penetration bias in X-band InSAR digital elevation models over the Greenland Ice Sheet.

GitHub Website
Example TanDEM-X scene showing interferometric coherence, penetration depth, and penetration bias

Developed as part of PhD research at DLR / ETH Zürich, combining physics-based scattering models with machine learning to correct systematic elevation bias caused by radar signal penetration into snow and ice, improving X-band InSAR DEM accuracy over the Greenland Ice Sheet.

Published as “Hybrid AI-Physical Modeling for Penetration Bias Correction in X-band InSAR DEMs: A Greenland Case Study” (with Georg Fischer, Ronny Hänsch, and Irena Hajnsek), extending an earlier conference paper presented at IGARSS 2024. An MLP predicts the parameters of a physical vertical scattering profile (the Exponential/Uniform-Volume model, and a more flexible Weibull profile) from InSAR observables — coherence, incidence angle, vertical wavenumber, interferometric phase, and backscatter — rather than regressing the penetration bias directly. The predicted parameters feed back into the physical model to compute the estimated bias, combining the interpretability and physical grounding of model-based inversion with the flexibility of a learned, data-driven fit.

Hybrid AI-physical pipeline: an MLP predicts scattering-profile parameters from InSAR observables, which feed into the physical coherence model to compute the penetration bias
Study area: footprints of the 18 TanDEM-X acquisitions and NASA IceBridge ATM flight tracks along a transect from Summit Camp to the Greenland east coast
Hybrid AI-physical pipeline (left), and the study area transect from Summit Camp to the Greenland east coast (right).

The model is trained and evaluated on 18 TanDEM-X acquisitions spanning a transect from the Greenland ice sheet summit to the east coast (winter 2017), using NASA IceBridge ATM LiDAR elevations as reference. To test generalization to real operational conditions, three training scenarios progressively withhold Height-of-Ambiguity (HoA) ranges: training on the full range, an interpolation test excluding mid-range HoA, and an extrapolation test excluding high HoA values.

Distribution of raw, uncorrected penetration bias across elevation bins
DEM error distributions before and after correction
Uncorrected penetration bias grows systematically with surface elevation (left); the hybrid Exponential model consistently performs best, RMSE as low as 0.52 m (R² = 0.94) on the full HoA range (right).
DEM error distributions for all nine modeling approaches, evaluated across all 18 TanDEM-X acquisitions
Across all nine model/scenario combinations, the hybrid Exponential model keeps its error distribution tightest and most centered near zero — the clearest picture of how the physical constraint pays off as training data gets less representative of the target conditions.

Code: github.com/IslamAlam/pydeepsar

References

  1. Islam Mansour, Georg Fischer, Ronny Hänsch, Irena Hajnsek Hybrid AI-physical Modeling for Penetration Bias Correction in X-band InSAR DEMs: A Greenland Case Study . 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , 2025 .
  2. Islam Mansour, Georg Fischer, Ronny Hänsch, Irena Hajnsek, Kostas Papathanassiou Correction of The Penetration Bias for InSAR DEM Via Synergetic AI-Physical Modeling: A Greenland Case Study . IGARSS 2024 – 2024 IEEE International Geoscience and Remote Sensing Symposium , 2024 .