Li et al. (2026) Assessing the global performance of a parsimonious soil temperature model for frozen ground prediction

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-03-19
• Authors: Donghui Li, Alexander Michalek, G. Villarini
• DOI: 10.1016/j.jhydrol.2026.135277

Research Groups

• High Meadow Environmental Institute, Princeton University, USA
• Department of Civil and Environmental Engineering, Princeton University, USA

Short Summary

This study globally evaluates a simplified soil temperature model for predicting frozen/unfrozen ground states using only air temperature and snow cover data. The model demonstrates robust global performance (average true frozen rate of 0.90, false frozen rate of 0.06), offering a computationally efficient solution for hydrological models, though it shows limitations in mountainous regions.

Objective

• To evaluate the global applicability and performance of a parsimonious soil temperature model (Rankinen et al., 2004) for predicting frozen/unfrozen ground conditions using only readily available air temperature and snow cover data.
• To assess the model's potential as a computationally efficient component for integrating frozen ground effects into large-scale distributed hydrological models, thereby improving flood prediction in cold regions.

Study Configuration

• Spatial Scale: Global (excluding Antarctica), with specific validation in the United States and Canada. ERA5-LAND reanalysis data at 9 km spatial resolution.
• Temporal Scale: Global assessment for the period 1980–2020. AmeriFlux validation using at least one year of continuous daily observations. All data aggregated to daily mean values.

Methodology and Data

• Models used: Simplified soil temperature prediction model by Rankinen et al. (2004). This physically-based model uses a reduced-form equation to derive soil surface temperature from air temperature, incorporating the insulating effect of snow cover via an empirical decay function based on snow depth.
• Data sources:
  • AmeriFlux network in-situ measurements: Daily mean soil temperature (2 cm below surface) from 167 sites across the United States and Canada. Co-located air temperature and snow depth from ERA5-LAND.
  • ERA5-LAND reanalysis data (Muñoz-Sabater et al., 2021): Hourly data aggregated to daily mean values for soil temperature (0–7 cm layer), 2-meter air temperature, and snow depth, covering the global land surface from 1980 to 2020.
• Model parameters: Five pre-calibrated parameter sets (for soil thermal conductivity KT, specific heat capacity of soil CS, specific heat capacity due to freezing and thawing CICE, and empirical snow parameter fS) from Rankinen et al. (2004) were applied uniformly.
• Evaluation metrics: True Frozen Rate (TFR) and False Frozen Rate (FFR) derived from contingency table analysis, and Receiver Operating Characteristic (ROC) curves. A soil temperature threshold of 0 °C was used to classify frozen/unfrozen conditions.

Main Results

• AmeriFlux validation: The model achieved an average true frozen rate (TFR) of 0.81 and an average false frozen rate (FFR) of 0.09 across 167 sites in the United States and Canada, demonstrating robust predictive capability.
• Global ERA5-LAND assessment (1980–2020):
  • Overall performance: The model showed excellent global capability, with an average TFR of 0.90 and a low FFR of 0.06. Performance remained high in temperate regions (TFR of 0.89).
  • ROC analysis: Approximately 60% of global grid cells achieved TFR above 0.9 and FFR below 0.1, confirming robust prediction ability.
  • Spatial limitations: Reduced accuracy (FFR exceeding 0.5 in some areas) was observed in mountainous regions (e.g., Tibetan Plateau, Rocky Mountains, Andes) and polar regions, attributed to the use of uniform parameters that do not capture complex snow dynamics and unique soil thermal properties in these terrains.
  • Seasonal variations: Monthly analyses revealed decreased accuracy during transitional warm months (April-July in Northern Hemisphere) due to lower TFRs, and during cold months (October-February) due to higher FFRs. These variations were largely attributed to the sensitivity of metrics to imbalanced frequencies of frozen/unfrozen events, rather than fundamental model limitations.

Contributions

• Provides the first comprehensive global assessment of the parsimonious soil temperature model by Rankinen et al. (2004), demonstrating its robust performance for frozen ground prediction using minimal inputs.
• Validates a computationally efficient and physically-based solution for incorporating frozen ground effects into large-scale hydrological models, addressing a critical gap in existing parsimonious models that often neglect snow cover insulation or require complex energy balance calculations.
• Identifies specific geographical limitations in mountainous and polar regions, highlighting the necessity for region-specific parameter calibration in complex terrains to improve accuracy.
• Offers significant implications for improved flood prediction in cold regions by enabling more accurate representation of reduced infiltration capacity due to frozen soil.

Funding

• Princeton University
• NOAA CIMES TASK III Project
• Next-Generation of NOAA water modeling: Climate Risks & Interactive Sub-seasonal to Seasonal PredictabilitY (CRISSPY) in the Earth System modeling framework

Citation

@article{Li2026Assessing,
  author = {Li, Donghui and Michalek, Alexander and Villarini, G.},
  title = {Assessing the global performance of a parsimonious soil temperature model for frozen ground prediction},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135277},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135277}
}