Chen et al. (2026) Spatial Heterogeneity and Drivers of Vertical Error in Global DEMs: An Explainable Machine Learning Approach in Complex Subtropical Coastal Zones

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Remote Sensing
• Year: 2026
• Date: 2026-04-10
• Authors: Junhui Chen, Fei Tang, Heshan Lin, Bo Huang, Xueping Lin
• DOI: 10.3390/rs18081125

Research Groups

Not explicitly mentioned in the paper text.

Short Summary

This study quantitatively decomposes the vertical errors of three 30 m global DEMs (COP30, NASADEM, and AW3D30) in Southeast China using ICESat-2 ATL08 data and an XGBoost-SHAP model, finding NASADEM has the lowest RMSE and identifying TRI, Land Cover, and specific sensor-related factors as dominant error drivers.

Objective

• To quantitatively decompose and understand the vertical errors of three 30 m global Digital Elevation Models (DEMs) – COP30, NASADEM, and AW3D30 – across the subtropical coastal region of Southeast China, using ICESat-2 ATL08 data as a reference and an eXtreme Gradient Boosting (XGBoost) model integrated with SHapley Additive exPlanations (SHAP).

Study Configuration

• Spatial Scale: Subtropical coastal region of Southeast China; global DEMs with 30 meter resolution.
• Temporal Scale: Not explicitly mentioned, but represents a static analysis of existing DEMs and ICESat-2 data.

Methodology and Data

• Models used: eXtreme Gradient Boosting (XGBoost) model, SHapley Additive exPlanations (SHAP).
• Data sources:
  • Digital Elevation Models (DEMs): COP30, NASADEM, AW3D30 (all 30 meter resolution).
  • Reference data: ICESat-2 ATL08 data.
  • Error drivers: Terrain Ruggedness Index (TRI), categorically encoded Land Cover, canopy height, topographic position.

Main Results

• NASADEM achieved the lowest Root Mean Square Error (RMSE) of 7.775 meters, followed by COP30 and AW3D30.
• The Terrain Ruggedness Index (TRI) and Land Cover were identified as universally dominant error drivers across all datasets.
• COP30 (X-band) is notably susceptible to canopy height, exhibiting significant positive bias in forests exceeding 15 meters.
• NASADEM (C-band) shows a systematic bias related to topographic position, typically overestimating ridges and underestimating valleys.
• AW3D30 (optical) is significantly affected by stereo-matching errors.
• A systematic error component of approximately 40% was quantified.

Contributions

• Provides a quantitative decomposition of vertical errors in global DEMs using high-accuracy ICESat-2 data and an explainable artificial intelligence (XGBoost-SHAP) approach.
• Identifies both universal (TRI, Land Cover) and sensor-specific error mechanisms for X-band (COP30), C-band (NASADEM), and optical (AW3D30) DEMs.
• Quantifies the systematic error component in these DEMs, offering insights into their inherent limitations.
• Offers data-driven recommendations for DEM selection and highlights priorities for accuracy improvements, specifically vegetation removal for radar DEMs and enhanced stereo-matching for optical models.

Funding

Not explicitly mentioned in the paper text.

Citation

@article{Chen2026Spatial,
  author = {Chen, Junhui and Tang, Fei and Lin, Heshan and Huang, Bo and Lin, Xueping},
  title = {Spatial Heterogeneity and Drivers of Vertical Error in Global DEMs: An Explainable Machine Learning Approach in Complex Subtropical Coastal Zones},
  journal = {Remote Sensing},
  year = {2026},
  doi = {10.3390/rs18081125},
  url = {https://doi.org/10.3390/rs18081125}
}