Zakeri et al. (2025) High-resolution snow water equivalent estimation: a data-driven method for localized downscaling of climate data

Identification

• Journal: Hydrology and earth system sciences
• Year: 2025
• Date: 2025-12-02
• Authors: Fatemeh Zakeri, Grégoire Mariethoz, Manuela Girotto
• DOI: 10.5194/hess-29-6935-2025

Research Groups

• Institute of Earth Surface Dynamics, Faculty of Geosciences and Environment, University of Lausanne, Lausanne, Switzerland
• Department of Environmental Science, Policy, and Management, University of California, Berkeley, Berkeley, CA, USA

Short Summary

This study develops a data-driven k-nearest neighbor method to downscale low-resolution climate data into daily high-resolution (500 m) snow water equivalent (SWE) estimates for mountainous regions. The approach successfully generates SWE data that closely matches reanalysis data, demonstrating that performance is highly dependent on the choice and accuracy of the climate model inputs.

Objective

• To develop a localized climate data downscaling method to estimate daily high-spatial-resolution (HR-SWE) data (500 m) from 1950 to the present, based on low-resolution climate predictors.
• To establish a statistical relationship between daily low-resolution climate data (temperature, precipitation, low-resolution SWE) and local reanalysis HR-SWE images using a k-nearest neighbor algorithm.

Study Configuration

• Spatial Scale:
  • Target SWE resolution: 500 m.
  • Study areas: California’s Sierra Nevada and Colorado’s Upper Colorado River Basin, Western United States.
  • Input climate data resolutions: 100 km (CMIP6) and 9 km (WRF-CMIP6).
• Temporal Scale:
  • Generated SWE data: Daily, from 1950 to present.
  • Reference SWE data (UCLA SWE): Water year 1984 to 2021 (daily).
  • Training period: 1985–2004 and 2011–2014.
  • Testing period: 2005–2010.
  • Climate variable intervals: Near interval (NI) of 4 days, Far interval (FI) of 60 days.

Methodology and Data

• Models used:
  • Core algorithm: k-nearest neighbor (k-NN) algorithm with a customized multivariate Manhattan distance metric.
  • Parameter optimization: Bayesian optimization for feature weights (α), sensitivity analysis for temporal intervals (NI, FI) and number of neighbors (K).
  • Climate models providing input predictors:
    • Coupled Model Intercomparison Project (CMIP) version 6 simulations (cnrm-esm2-1, ec-earth3-veg, mpi-esm1-2) at 100 km resolution.
    • Dynamically downscaled CMIP6 over the Western United States using the Weather Research and Forecasting (WRF) model (WRF-CMIP6) (cnrm-esm2-1, ec-earth3-veg, mpi-esm1-2) at 9 km resolution.
• Data sources:
  • Reference HR-SWE (training and validation): UCLA SWE dataset (reanalyzed SWE for the Western United States, 16 arcsec / ~500 m resolution, 1984–2021), derived from Landsat-based fractional snow-covered area observations and Bayesian data assimilation.
  • Input climate variables (predictors): Daily minimum/maximum air temperature, total precipitation, and surface downwelling shortwave radiation from CMIP6 and WRF-CMIP6.
  • Validation datasets:
    • Snow Data Assimilation System (SNODAS) 1 km SWE product.
    • Daymet V4 1 km SWE product.
    • University of Arizona (UA) 4 km SWE product.
  • Ancillary data: SNOTEL network (for in situ locations), MODIS Land Cover Type (MCD12Q1) Version 6.1, NASA SRTM Digital Elevation 30 m data, CONUS Drought Indices dataset (SPEI).

Main Results

• The proposed data-driven k-NN method effectively downscales low-resolution climate data to daily 500 m SWE, with outputs closely matching reanalysis data in statistical properties.
• The accuracy of downscaled SWE is highly sensitive to the choice and accuracy of the climate model inputs (e.g., precipitation and temperature data).
• In Colorado, the cnrm-esm2-1 model demonstrated superior accuracy at both 100 km and 9 km resolutions (e.g., mean RMSE of 0.07 m for cross-validation at 9 km).
• In California, the ec-earth3-veg model achieved the best performance, particularly at 9 km resolution (e.g., mean RMSE of 0.13 m for cross-validation at 9 km).
• Climate inputs at 9 km resolution generally provided slightly better accuracy than 100 km resolution, though the difference was modest and varied by region and model.
• Models generally performed better or comparably at higher elevations (>3000 m) compared to medium elevations (2000 m to 3000 m).
• The method successfully captures fine-scale SWE patterns in complex terrains and can reconstruct historical SWE values for periods with only low-resolution climate data.

Contributions

• Introduces a novel data-driven k-NN method for daily SWE downscaling to 500 m resolution using low-resolution climate data, eliminating the need for high-resolution meteorological inputs or extensive ground observations.
• Enhances traditional k-NN downscaling by incorporating "far" (60 days) and "near" (4 days) temporal intervals of climate data, improving its ability to handle dynamic variables and preserve extreme events across broader temporal ranges.
• Demonstrates the potential to generate high-resolution SWE data for historical periods (1950 to present) where such data were previously unavailable, and for applications in data-scarce regions.
• Provides a computationally efficient alternative to complex physical snow models for generating high-resolution SWE over large spatial and temporal domains.
• Highlights that climate model selection and input quality are crucial drivers for SWE estimation performance, sometimes more so than merely increasing input resolution.

Funding

• Université de Lausanne (UNIL) Mobility Fellowship (MD0012)
• Swiss National Science Foundation (SNSF) project "Deep-time Synthetic Data Cubes to Enable Long-term Hydrological Modeling" (200021_204130)

Citation

@article{Zakeri2025Highresolution,
  author = {Zakeri, Fatemeh and Mariethoz, Grégoire and Girotto, Manuela},
  title = {High-resolution snow water equivalent estimation: a data-driven method for localized downscaling of climate data},
  journal = {Hydrology and earth system sciences},
  year = {2025},
  doi = {10.5194/hess-29-6935-2025},
  url = {https://doi.org/10.5194/hess-29-6935-2025}
}