Explicit Features Versus Implicit Spatial Relations in Geomorphometry: A Comparative Analysis for DEM Error...

Area of Science:

• Geospatial analysis
• Machine learning in Earth sciences
• Digital Elevation Model (DEM) refinement

Background:

• Global DEMs have biases due to acquisition geometry and surface penetration.
• Sparse altimetry data presents challenges for accurate terrain modeling.
• Evaluating advanced geometric deep learning against traditional methods is crucial.

Purpose of the Study:

• To compare the performance of Graph Neural Networks (GNNs) against XGBoost for DEM accuracy improvement.
• To assess if complex geometric deep learning justifies performance gains over feature engineering with sparse altimetry data.
• To investigate the underlying mechanisms driving model performance differences.

Main Methods:

• Comparative framework using four DEM products (ALOS World 3D, Copernicus DEM, SRTM GL1, TanDEM-X).
• Utilized Sichuan Province, China, as a study area with sparse altimetry supervision.
• Evaluated Hybrid GNN models against a systematically optimized XGBoost baseline.

Main Results:

• Hybrid GNN models showed no statistically significant accuracy advantage over XGBoost (RMSE parity).
• A scale mismatch between altimetry footprint spacing and terrain resolution was identified.
• XGBoost relied on deterministic features (terrain aspect, vegetation density), while GNNs exhibited attribution stochasticity.

Conclusions:

• For geospatial regression with sparse supervision, domain-knowledge-based physical features are superior to complex "black box" architectures.
• "Physics Trumps Geometry" in scenarios with limited data.
• Prioritizing explanatory stability over marginal accuracy gains is essential for trusted Geo-AI.