González-Abad et al. (2026) Pre-training for deep statistical climate downscaling: enhancing consistency and robustness across regional datasets

Identification

• Journal: Geoscientific model development
• Year: 2026
• Date: 2026-07-01
• Authors: José González-Abad, Maialen Iturbide, Alfonso Hernanz, José Manuel Gutiérrez
• DOI: 10.5194/gmd-19-5781-2026

Research Groups

• Instituto de Física de Cantabria (IFCA), CSIC-Universidad de Cantabria, Santander, Spain
• Spanish Meteorological Agency (AEMET), Madrid, Spain

Short Summary

This study investigates the use of pre-training in deep statistical climate downscaling to enhance consistency and robustness across regional datasets. The authors evaluate the effectiveness of different fine-tuning strategies to adapt a pre-trained model to alternative national and regional station datasets.

Objective

• Investigate the benefits of pre-training for deep statistical downscaling, particularly in cases with limited spatial and/or temporal data availability.
• Develop cost-effective downscaling approaches that preserve coherence with national-scale results while accommodating local dataset characteristics.

Study Configuration

• Spatial Scale: Peninsular Spain (36–44°N, 9.5°W–3.5°E)
• Temporal Scale: Daily minimum and maximum temperature and precipitation for the period 1980–2021

Methodology and Data

• Models used: DeepESD architecture with a single kernel in the final convolutional layer
• Data sources:
  • ERA5 reanalysis dataset (predictors)
  • ROCIO-IBEB gridded dataset (predictand for pre-training)
  • STATIONS-IBEB and STATIONS-CAT station datasets (predictands for fine-tuning)

Main Results

• Fine-tuning a pre-trained model on alternative national or regional datasets improves the consistency and robustness of climate change signals.
• The pre-trained model provides a robust foundation for statistical downscaling, reducing epistemic uncertainty and improving the reliability of future climate projections.

Contributions

• This study represents a step toward standardizing DL-based downscaling models to ensure more coherent and consistent climate projections across national and regional scales.
• The proposed approach can be particularly valuable in low-data regimes, where limited spatial and/or temporal data availability is common.

Funding

• Not specified

Citation

@article{GonzálezAbad2026Pretraining,
  author = {González-Abad, José and Iturbide, Maialen and Hernanz, Alfonso and Gutiérrez, José Manuel},
  title = {Pre-training for deep statistical climate downscaling: enhancing consistency and robustness across regional datasets},
  journal = {Geoscientific model development},
  year = {2026},
  doi = {10.5194/gmd-19-5781-2026},
  url = {https://doi.org/10.5194/gmd-19-5781-2026}
}