Hadjipetrou (2026) A review of statistical methods for climate downscaling: the underexplored potential of geostatistical simulation

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2026
• Date: 2026-03-15
• Authors: Stylianos Hadjipetrou
• DOI: 10.1007/s00704-026-06120-2

Research Groups

• Department of Civil Engineering and Geomatics, Cyprus University of Technology, Limassol 3036, Cyprus

Short Summary

This review synthesizes developments in statistical and stochastic climate downscaling, critically assessing various methods including regression, weather generators, analogs, and machine learning. It highlights the significant, yet underexplored, potential of geostatistical simulation, particularly Multiple-Point Statistics, to provide spatially coherent and uncertainty-aware fine-scale climate information.

Objective

• To provide a critical overview of existing statistical downscaling methods, assessing their strengths and limitations, in light of recent advances in machine learning for generating high-resolution climate information.
• To highlight the untapped potential of geostatistical simulation for modeling the relationship between large-scale and local climate information, and to advocate for its broader integration into downscaling frameworks.

Study Configuration

• Spatial Scale: Review of methods downscaling from coarse-resolution Global Climate Models (GCMs) (e.g., 25 km, 1 degree) to finer regional, local, or observation scales (e.g., 2 km, 1 km, site-specific).
• Temporal Scale: Review of methods downscaling from monthly GCM outputs to daily or hourly data, often utilizing historical periods (e.g., decades to 180 years) for training and validation.

Methodology and Data

• Models used:
  • Empirical/Statistical Downscaling (ESD): Regression-based methods (Multiple Linear Regression, Generalized Linear Model, Principal Component Regression, Canonical Correlation Analysis, Bias Correction and Spatial Disaggregation, Change Factor), Weather Generators (e.g., Statistical Downscaling Model, Markov chain processes), Weather Typing/Analogs (Weather Typing, Constructed Analogs, Analog Method).
  • Machine Learning (ML) and Deep Learning (DL): Artificial Neural Networks, Genetic Programming, Support Vector Machine, Relevance Vector Machine, Convolutional Neural Networks, Generative Adversarial Networks, Diffusion Models.
  • Geostatistical Methods: Kriging variants (Regression Kriging, Downscaling Cokriging, Kriging with External Drift, Area-to-Point Regression Kriging), Sequential Gaussian Simulation, Sequential Indicator Simulation, Multiple-Point Statistics (Direct Sampling, Quick Sampling).
  • Disaggregation Models: Multiplicative cascade models, Gibbs Sampling Disaggregation Model.
• Data sources:
  • Global Climate Model (GCM) and Regional Climate Model (RCM) outputs.
  • Historical observations (station data, gridded observational datasets).
  • Satellite-derived data (e.g., Landsat 7 imagery, Sentinel-1, CMORPH, radar reflectivity).
  • Reanalysis data (e.g., ERA5, UERRA–HARMONIE).
  • Training images (for Multiple-Point Statistics).
  • Ancillary data (e.g., topography).

Main Results

• Statistical downscaling methods are computationally efficient and flexible but are often limited by the stationarity assumption of predictor-predictand relationships and the need for extensive historical data.
• Traditional regression-based methods offer interpretability but tend to smooth out spatial and temporal variability, potentially underestimating local extremes.
• Machine learning and deep learning approaches excel at modeling complex, nonlinear relationships and capturing extremes, but demand large, high-quality training datasets, incur high computational costs, and face challenges in generalization and interpretability.
• Geostatistical simulation techniques, particularly Multiple-Point Statistics (MPS), are significantly underutilized in climate downscaling despite their proven ability to capture complex spatial dependencies, preserve spatial continuity, reproduce fine-scale patterns, and quantify uncertainty through multiple equiprobable realizations.
• Traditional geostatistical methods (e.g., Kriging, Sequential Gaussian Simulation) often rely on two-point statistics, which limits their capacity to represent higher-order, non-linear spatial patterns and can lead to smoothing effects.
• The review advocates for hybrid methodologies that integrate geostatistical simulation with advanced machine learning techniques (e.g., for automated training image selection or synthesis) to generate more robust, fine-scale, and uncertainty-aware climate information.

Contributions

• Provides a comprehensive and critical review of the state-of-the-art in statistical and stochastic climate downscaling, including recent advancements in machine learning.
• Identifies and highlights the significant, yet underexplored, potential of geostatistical simulation, especially Multiple-Point Statistics, for improving climate downscaling.
• Advocates for the broader integration of geostatistical simulation into downscaling frameworks to enhance spatial coherence, structural realism, and the quantification of uncertainty in climate projections.
• Outlines future research directions, emphasizing the synergistic potential of combining geostatistical simulation with advanced machine learning and artificial intelligence techniques.

Funding

• Open access funding provided by the Cyprus Libraries Consortium (CLC).
• Open Access funding was provided through the Cyprus Libraries Consortium (CLC) Open Access agreement with Springer Nature.

Citation

@article{Hadjipetrou2026review,
  author = {Hadjipetrou, Stylianos},
  title = {A review of statistical methods for climate downscaling: the underexplored potential of geostatistical simulation},
  journal = {Theoretical and Applied Climatology},
  year = {2026},
  doi = {10.1007/s00704-026-06120-2},
  url = {https://doi.org/10.1007/s00704-026-06120-2}
}