Wen et al. (2025) Spatial distribution and variation trends of soil freezing front on the Qingzang Plateau revealed by machine learning models

Identification

• Journal: Climate Dynamics
• Year: 2025
• Date: 2025-12-11
• Authors: Bo Wen, Tingbin Zhang, Xiaobing Zhou, Guihua Yi, Bin Yang, Dongmei Chen, Jingji Li, Xiang Wang, Xianglong Ma
• DOI: 10.1007/s00382-025-07989-x

Research Groups

• College of Earth and Planetary Sciences, Chengdu University of Technology, Chengdu, China
• Middle Yarlung Zangbo River Natural Resources Observation and Research Station of Tibet Autonomous Region, Lhasa, China
• State Environmental Protection Key Laboratory of Synergetic Control and Joint Remediation for Soil and Water Pollution, Chengdu University of Technology, Chengdu, China
• Geophysical Engineering Department, Montana Tech of the University of Montana, Butte, MT, USA
• College of Geography and Planning, Chengdu University of Technology, Chengdu, China
• Research Center of Applied Geology of China Geological Survey, Chengdu, China
• Department of Geography and Planning, Queen’s University, Kingston, ON, Canada
• College of Ecological Environment, Chengdu University of Technology, Chengdu, China
• Key Laboratory of Ecosystem Network Observation and Modeling, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing, China

Short Summary

This study simulated the spatiotemporal variations of soil freezing front depth (SFFD) on the Qingzang Plateau (QP) from 1982 to 2019 using machine learning models, revealing a significant and accelerating shallowing trend across both permafrost and seasonally frozen ground regions.

Objective

• To simulate the soil freezing front depth (SFFD) using machine learning techniques.
• To compare the performance of Random Forest (RF), Support Vector Machine (SVM), and Multilayer Perceptron (MLP) to identify the optimal model for predicting SFFD.
• To characterize the spatiotemporal variation patterns of the SFFD during the freezing period on the Qingzang Plateau.

Study Configuration

• Spatial Scale: Qingzang Plateau (QP), covering approximately 2.61 × 10^6 square kilometers, with analysis distinguishing between permafrost and seasonally frozen ground regions.
• Temporal Scale: 1982–2019 for SFFD simulation and trend analysis, focusing on the freezing period from October to January. Training data covered 1982–2008, and an independent test set covered 2009–2015.

Methodology and Data

• Models used: Random Forest (RF), Support Vector Machine (SVM), and Multilayer Perceptron (MLP). The Random Forest model was identified as optimal.
• Data sources:
  • Monthly soil freezing depth observations (1982–2015) from 111 meteorological stations on the QP, measured by the China Meteorological Administration (CMA).
  • Climate variables: Monthly mean air temperature, monthly mean minimum air temperature (from ChinaClim_Time-series dataset), degree days of freezing (calculated from temperature), monthly precipitation, monthly mean specific humidity, monthly mean wind speed, monthly mean pressure, monthly mean shortwave radiation, and monthly mean longwave radiation (from High-resolution Near-surface Meteorological Forcing Dataset for the Third Pole Region).
  • Soil parameters: Soil saturated water content, thermal conductivity of unfrozen saturated soils, thermal conductivity of frozen saturated soils, thermal conductivity of dry soils (from Global High-resolution Dataset of Soil Hydraulic and Thermal Parameters for Land Surface Modeling), soil organic matter, bulk density, silt content, clay content, and sand content (from Soil Database of China for Land Surface Modeling).
  • Gridded climate variables were interpolated to a uniform resolution of 0.01° × 0.01°.

Main Results

• The Random Forest (RF) model demonstrated superior performance in simulating SFFD (R² = 0.81, RMSE = 28.09 cm, MAE = 18.02 cm). Model Scenario 1 (excluding previous month's freezing depth as a predictor) showed the strongest spatial generalization capability and robustness in out-of-sample prediction (R² = 0.93, RMSE = 17.00 cm, MAE = 12.32 cm on the independent test set).
• Degree Days of Freezing (DDF) was the dominant predictor for SFFD, followed by Soil Organic Matter (SOM). Lower temperatures and less SOM were associated with deeper freezing fronts.
• The average SFFD on the QP increased from 28.40 cm in October to 119.70 cm in January (1982–2019). Spatially, SFFD was deeper in the west and north and shallower in the east and south.
• A consistent and statistically significant shallowing trend in SFFD was observed across the QP from 1982 to 2019. The monthly shallowing rates were -0.11 cm/yr in October, -0.22 cm/yr in November, -0.29 cm/yr in December, and -0.38 cm/yr in January.
• Both permafrost and seasonally frozen ground regions exhibited significant declines in SFFD. In permafrost areas, shallowing rates were relatively stable throughout the freezing period, while in seasonally frozen regions, the shallowing trend progressively intensified from October to January.
• The rate of SFFD shallowing accelerated markedly when comparing the sub-periods 1983–2000 and 2001–2019 (e.g., November shallowing rate in permafrost increased by 85%, and January in seasonally frozen regions by 91%).
• The influence of elevation on SFFD trends exhibited seasonal variability: weak in October, strengthening to a predominantly negative influence in November–December, and becoming nonlinear in January with the strongest negative impact at mid-high elevations (4000–4500 m).

Contributions

• This study is the first to employ machine learning algorithms (Random Forest, Support Vector Machine, Multilayer Perceptron) to simulate monthly soil freezing front depth (SFFD) on the Qingzang Plateau.
• It provides a comprehensive analysis of the spatiotemporal dynamics and long-term (1982–2019) variation trends of SFFD, addressing a gap in previous research primarily focused on annual maximum freezing depth.
• The research identified and validated an optimal machine learning model (Random Forest without previous month's SFFD) for robust and accurate out-of-sample prediction of SFFD.
• It quantified the accelerating shallowing trends of SFFD across both permafrost and seasonally frozen ground regions, highlighting distinct regional responses and the influence of elevation.
• The study demonstrated the scalability of the monthly-trained RF model to accurately estimate both annual maximum freeze depths and daily freezing front depths, expanding its applicability.

Funding

• Open Project of Middle Yarlung Zangbo River Natural Resources Observation and Research Station (Grant No. 2024YJZKF002)
• Second Tibetan Plateau Scientific Expedition and Research Program (Grant No. 2019QZKK0307)

Citation

@article{Wen2025Spatial,
  author = {Wen, Bo and Zhang, Tingbin and Zhou, Xiaobing and Yi, Guihua and Yang, Bin and Chen, Dongmei and Li, Jingji and Wang, Xiang and Ma, Xianglong},
  title = {Spatial distribution and variation trends of soil freezing front on the Qingzang Plateau revealed by machine learning models},
  journal = {Climate Dynamics},
  year = {2025},
  doi = {10.1007/s00382-025-07989-x},
  url = {https://doi.org/10.1007/s00382-025-07989-x}
}