Hybrid AI-Physical Modeling for Penetration Bias Correction in X-band InSAR DEMs: A Greenland Case Study

TL;DR

Addresses penetration bias in InSAR DEMs caused by radar signal penetration into snow and ice. Proposes a Hybrid AI–Physical Modeling framework that uses parametric vertical scattering profiles (Exponential and Weibull) and an MLP to predict profile parameters from InSAR features, assessed on Greenland TanDEM-X data under three HoA-based training scenarios. The Exponential hybrid model achieves the best accuracy (RMSE ≈ $0.52$ m, $R^2 ≈ 0.94$) and shows strong generalization when acquisition diversity is limited, outperforming a pure ML baseline. The work demonstrates robustness, interpretability, and the potential to fuse multi-sensor data and extend to other missions and wavelengths.

Abstract

Digital elevation models derived from Interferometric Synthetic Aperture Radar (InSAR) data over glacial and snow-covered regions often exhibit systematic elevation errors, commonly termed "penetration bias." We leverage existing physics-based models and propose an integrated correction framework that combines parametric physical modeling with machine learning. We evaluate the approach across three distinct training scenarios - each defined by a different set of acquisition parameters - to assess overall performance and the model's ability to generalize. Our experiments on Greenland's ice sheet using TanDEM-X data show that the proposed hybrid model corrections significantly reduce the mean and standard deviation of DEM errors compared to a purely physical modeling baseline. The hybrid framework also achieves significantly improved generalization than a pure ML approach when trained on data with limited diversity in acquisition parameters.

Figures (18)

• Figure 1: Qualitative example from a representative region. Left: Interferometric coherence (range: 0.3--1). Center: One-way penetration depth ($d_{\mathrm{pen}}$, in meters) predicted by the MLP component of our hybrid framework, which then feeds into the Exponential profile. Right: Final penetration bias ($p_{\mathrm{bias}}$, in meters). This hybrid approach leverages InSAR features (e.g., coherence) to predict $d_{\mathrm{pen}}$, which is subsequently used to estimate $p_{\mathrm{bias}}$ for correcting DEM elevations over ice and snow.
• Figure 3: Overview of our hybrid modeling pipeline. An MLP predicts scattering profile parameters that feed into the physical model for computing the estimated bias $\hat{p}_{\mathrm{bias}}$. We use MSE loss against a reference bias (e.g., LiDAR).
• Figure 4: Overview of TanDEM-X scenes (blue) and ATM flight tracks (red) over the study area in Greenland. The Summit Camp location is marked with a yellow star.
• Figure 5: Overview of the mosaicked dataset used in our study. Each panel shows a different attribute for the 2017 TanDEM-X acquisitions over Greenland.
• Figure 6: Bias distribution across elevation bins, computed using ATM LiDAR as reference. The penetration bias increases at higher elevations due to smaller scattering losses from less melt-refreeze features in the subsurface.