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

Technical Summary: Hybrid AI–Physical Modeling for Penetration Bias Correction in X-band InSAR DEMs

Problem Statement
Digital Elevation Models (DEMs) derived from Interferometric Synthetic Aperture Radar (InSAR) over glacial and snow-covered regions suffer from systematic elevation errors known as "penetration bias." This phenomenon occurs because X-band radar signals penetrate dry snow and ice, causing the measured phase center to lie below the actual surface. As demonstrated by comparisons between ICESat-2 and TanDEM-X data, this bias can reach several meters and varies significantly with terrain elevation and snow properties. While purely physical models (e.g., Uniform Volume) offer interpretability without training data, they often fail to capture complex, real-world scattering scenarios. Conversely, purely data-driven Machine Learning (ML) approaches can model local variations but often lack robustness and generalization capabilities, particularly when training data lacks diversity in acquisition geometries.

Methodology
The authors propose a hybrid framework that synergistically combines parametric physical modeling with a Machine Learning component (a Multi-Layer Perceptron, or MLP). The core objective is to estimate the penetration bias ($p_{bias}$) by predicting the parameters of a vertical scattering profile $f(z)$ from InSAR observables.

1. Physical Models: The study evaluates two parametric profiles governing the vertical scattering of signals:
   • Exponential Profile: Derived from the Uniform Volume (UV) model, assuming a constant signal extinction coefficient. It requires estimating a single parameter: the one-way penetration depth ($d_{pen}$).
   • Weibull Profile: A more flexible model capable of capturing changes in signal extinction due to grain size and density variations with depth. It requires estimating two parameters: a scale parameter ($\lambda_w$) and a shape parameter ($k_w$).
2. Hybrid Architecture: Instead of assuming fixed physical parameters, an MLP is trained to predict the specific parameters ($\hat{d}_{pen}$, $\hat{\lambda}_w$, $\hat{k}_w$) for each pixel based on input features such as volumetric decorrelation, incidence angle, vertical wavenumber, interferometric phase, and backscatter. These predicted parameters are then fed into the physical equations (specifically the volume decorrelation equation) to compute the final estimated bias.
3. Training Scenarios: To rigorously assess generalization, the authors define three training scenarios based on the Height of Ambiguity (HoA), a key acquisition geometry parameter:
   • All: Training on the full range of available HoA values.
   • Interpolation: Training excludes mid-range HoA values (50–60 m) to test interpolation capabilities.
   • Extrapolation: Training excludes high HoA values (>70 m) to test extrapolation performance.

Key Contributions

• Hybrid Framework: The paper presents an integrated correction framework that embeds physical constraints (Exponential and Weibull profiles) into a data-driven learning process, allowing the model to learn scattering profile parameters directly from InSAR data.
• Systematic Generalization Analysis: Unlike previous works that may focus solely on accuracy, this study systematically evaluates model robustness across different acquisition geometries (HoA) by explicitly testing interpolation and extrapolation capabilities.
• Comparative Evaluation: The study provides a direct comparison between the hybrid approach, a purely physical baseline (UV model), and a purely data-driven MLP baseline.

Results
Experiments conducted on TanDEM-X data over the Greenland ice sheet, validated against NASA IceBridge ATM LiDAR data, yield the following findings:

• Baseline Performance: The uncorrected DEM exhibits a large negative bias (mean $\approx -5.14$ m). The purely physical UV model reduces this bias but results in high RMSE (2.07 m) and low $R^2$ (0.06), indicating it is insufficient for capturing actual scattering behavior.
• Hybrid vs. Pure ML: When trained on the full range of HoA values, both the hybrid models and the pure MLP achieve comparable high accuracy (e.g., Exponential hybrid RMSE $\approx 0.52$ m, $R^2 \approx 0.94$).
• Generalization Robustness: The critical advantage of the hybrid approach emerges when training data is limited in diversity:
  • Interpolation: The Exponential hybrid model maintains low error and high $R^2$ (0.94) even when mid-range HoA values are excluded, outperforming the Weibull and MLP models which show increased variability.
  • Extrapolation: The MLP model suffers a significant performance drop (RMSE increases to 1.27 m, $R^2$ drops to 0.64) when tested on unseen high HoA values. In contrast, the Exponential hybrid model remains robust (RMSE 0.88 m, $R^2$ 0.83).
• Model Complexity: The simpler Exponential model (one parameter) proves more robust and stable than the Weibull model (two parameters), which shows moderate degradation under extrapolation, likely due to the ill-posed nature of estimating two parameters without sufficient constraints.

Significance and Claims
The paper claims that embedding physical constraints into the learning process significantly enhances model performance and stability, particularly in operational scenarios where high-quality reference data or diverse acquisition geometries are limited. The authors assert that:

1. The Exponential hybrid model is the most effective method for correcting DEM penetration bias, offering the lowest errors and highest stability across all tested scenarios.
2. Incorporating physical constraints reduces the dependency on highly diverse training datasets, enabling better generalization to unseen acquisition geometries (HoA conditions).
3. The framework improves interpretability by grounding predictions in physical scattering profiles, making it a robust solution for bias correction that can potentially be adapted to other satellite missions (e.g., Sentinel, Harmony) operating at different wavelengths.

The study concludes that while pure ML methods are effective with balanced data, the hybrid approach is superior for ensuring reliability when data diversity is constrained, a common challenge in large-scale InSAR applications.