Abstract
The purpose of this research is to study how different machine learning and statistical models can be used to predict bedrock topography under the Greenland ice sheet using ice-penetrating radar and satellite imagery data. Accurate bed topography representations are crucial for understanding ice sheet stability and vulnerability to climate change. We explore nine predictive models including dense neural network, long-short term memory, variational auto-encoder, extreme gradient boosting (XGBoost), gaussian process regression, and kriging based residual learning. Model performance is evaluated with mean absolute error (MAE), root mean squared error (RMSE), coefficient of determination ($R^2$), and terrain ruggedness index (TRI). In addition to testing various models, different interpolation methods, including nearest neighbor, bilinear, and kriging, are also applied in preprocessing. The XGBoost model with kriging interpolation exhibit strong predictive capabilities but demands extensive resources. Alternatively, the XGBoost model with bilinear interpolation shows robust predictive capabilities and requires fewer resources. These models effectively capture the complexity of the terrain hidden under the Greenland ice sheet with precision and efficiency, making them valuable tools for representing spatial patterns in diverse landscapes.
Introduction
Predicting the response of the Greenland ice sheet to climate change is a critical component of sea-level rise projections. Central to this prediction is the subglacial bed topography—the "shape" of the land beneath the ice. This topography dictates ice flow dynamics; for instance, a retrograde slope (where the bed deepens inland) can lead to unstable, runaway glacial retreat.
While airborne ice-penetrating radar provides high-accuracy measurements, it is limited to the narrow tracks directly beneath flight paths, leaving vast areas unobserved. Traditionally, physics-based mass conservation models like BedMachine have been used to interpolate these gaps. However, these models often result in overly smoothed terrain that fails to capture the rugged complexity of subglacial fjords. This study investigates whether machine learning can leverage abundant satellite-derived surface data to infer these hidden landscapes with greater precision.
Methodology and Datasets
The study focused on a 32,400 km² region within the Upernavik glacier system in West Greenland. The researchers evaluated nine predictive approaches, including statistical methods (Gaussian Process Regression, Universal Kriging), deep learning (MLP, LSTM, VAE), and gradient boosting (XGBoost).
Data Integration
The models were trained using a suite of surface variables as input features:
- Surface Elevation: Derived from the Greenland Ice Mapping Project (GIMP).
- Ice Flow Velocity: Sourced from Landsat-8 and Sentinel-1.
- Thinning Rates: Measured via ICESat-2.
- Surface Mass Balance: Provided by the RACMO model.
The target variable (ground truth) consisted of subglacial bed elevation measurements from NASA’s Operation IceBridge. A key technical challenge involved "interpolation ablation"—testing how different preprocessing methods (Nearest Neighbor, Bilinear, and Kriging) for aligning radar tracks with surface grids affected final model accuracy.
Key Results and Discussion
The Superiority of XGBoost
XGBoost emerged as the top-performing model, significantly outperforming the physics-based BedMachine baseline. While BedMachine achieved an $R^2$ of 0.842 and an RMSE of 71.55m, the XGBoost model reached an $R^2$ of 0.967 with an RMSE of only 32.68m. Beyond standard error metrics, the researchers utilized the Terrain Ruggedness Index (TRI) to assess qualitative fidelity. While physics models produced a TRI of ~3.3 (indicating excessive smoothing), XGBoost produced a TRI of ~14-15, much closer to the observed ruggedness of glacial terrain.
The Kriging Paradox
An interesting finding was the performance of Universal Kriging. While it achieved the lowest Mean Absolute Error (MAE), it suffered from poor RMSE and $R^2$ values. This discrepancy occurs because Kriging is highly effective at leveraging local spatial correlations but fails in data-sparse regions, leading to extreme outliers that skew squared-error metrics.
Preprocessing Efficiency
While Kriging interpolation during preprocessing yielded the highest accuracy, it was computationally prohibitive for large-scale applications. The study found that Bilinear interpolation offered a "Goldilocks" solution—providing robust predictive power with significantly lower resource requirements, making it the preferred choice for continental-scale mapping.
Conclusion
This research demonstrates that machine learning models, particularly XGBoost, can effectively bridge the gap between sparse radar observations and the need for high-resolution subglacial maps. By capturing the jagged reality of subglacial fjords that traditional models smooth over, these methods provide a more accurate foundation for modeling ice sheet stability. Future work will explore using the Terrain Ruggedness Index as a training constraint to further refine the visual and physical accuracy of these hidden landscapes.
Summary updated