Dumitrescu (2026) A deep learning framework for gridding daily climate variables from a sparse station network

Identification

• Journal: Geoscientific model development
• Year: 2026
• Date: 2026-09-21
• Authors: Alexandru Dumitrescu
• DOI: 10.5194/gmd-19-8895-2026

Research Groups

• Department of Climatology, Meteo Romania (National Meteorological Administration)
• University of Bucharest

Short Summary

This study evaluates a deep learning framework, the Spatial Multi-Attention Conditional Neural Process (SMACNP), for gridding daily climate variables from a sparse station network in Romania. The SMACNP model outperforms Regression Kriging (RK) for both temperature and precipitation.

Objective

• Evaluate the performance of SMACNP for interpolating air temperature and precipitation from a sparse station network.
• Compare SMACNP with Regression Kriging (RK) as a geostatistical baseline.

Study Configuration

• Spatial Scale: Romania, a region characterized by complex topography including the Carpathian Mountains, plateaus, and plains.
• Temporal Scale: Daily data from 2020 to 2023.

Methodology and Data

• Models used: SMACNP (Spatial Multi-Attention Conditional Neural Process) and Regression Kriging (RK)
• Data sources: Meteorological station data and a topographic variable extracted from high-resolution Digital Elevation Model (DEM)

Main Results

• The SMACNP model outperforms RK for both temperature and precipitation, with the localized encoder variant achieving the best performance.
• The SMACNP (Localized) model achieves the lowest Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) for both variables.

Contributions

• This study demonstrates the effectiveness of SMACNP for interpolating climate data from a sparse station network in complex terrain.
• The results highlight the importance of using localized attention mechanisms to capture spatial dependencies in data-scarce regimes.

Funding

• This research was funded by the National Meteorological Administration, Romania.

Citation

@article{Dumitrescu2026deep,
  author = {Dumitrescu, Alexandru},
  title = {A deep learning framework for gridding daily climate variables from a sparse station network},
  journal = {Geoscientific model development},
  year = {2026},
  doi = {10.5194/gmd-19-8895-2026},
  url = {https://doi.org/10.5194/gmd-19-8895-2026}
}