Su et al. (2026) Prediction of spring agricultural drought in cold and arid regions on the basis of soil freeze thaw processes

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-03-31
• Authors: Yi Su, Wei Pei, Shaoting Liu
• DOI: 10.1016/j.ejrh.2026.103382

Research Groups

• College of Arts and Sciences, Northeast Agricultural University, Harbin, China
• School of Water Conservancy & Civil Engineering, Northeast Agricultural University, Harbin, China
• Department of Information Engineering, Hebei Institute of Mechanical and Electrical Technology, Xingtai, China

Short Summary

This study developed and evaluated a spring agricultural drought prediction method for cold and arid regions, based on soil freeze-thaw processes, finding that the Projection Pursuit Regression (PPR) model significantly outperformed other machine learning models, particularly in shallow soil layers and arid conditions.

Objective

• To develop and evaluate a spring agricultural drought prediction method for cold and arid regions, specifically Qiqihar city, by incorporating soil freeze-thaw processes and utilizing a Projection Pursuit Regression (PPR) model with the Standardized Soil Moisture Index (SSMI) to enhance prediction accuracy and reduce drought uncertainty.

Study Configuration

• Spatial Scale: Qiqihar city, Heilongjiang Province, China (42,469 square kilometers), spanning 46.25°N to 48.25°N latitude and 122.75°E to 126.25°E longitude, with 22 sampling points at 0.5° intervals. The study area was also divided into northern, central, and southern subregions based on latitude.
• Temporal Scale: 41 years, from 1980 to 2020. A rolling annual training-validation approach was used, with training sets from 1980-2016, 1980-2017, 1980-2018, and 1980-2019 to predict for 2017, 2018, 2019, and 2020, respectively. The prespring period was divided into four phases: prefreeze, rapid freezing, stable freezing, and thawing.

Methodology and Data

• Models used: Projection Pursuit Regression (PPR) model, Principal Component Regression (PCR) model, Partial Least Squares (PLS) model, Projection Pursuit Regression coupled Partial Least Squares (PLS-PPR) model, Long Short-Term Memory (LSTM) model, and Random Forest (RF) model. The PPR model utilized an Ant Colony Optimization (ACO) algorithm for parameter optimization.
• Data sources: China Meteorological Administration's Global Atmosphere/Land Surface Reanalysis Product (CMA-RA) from the China Meteorological Data Network (http://data.cma.cn/). Data included 48 explanatory variables related to spring drought (e.g., total precipitation, potential evapotranspiration, soil moisture content at 10 cm, 40 cm, 100 cm, 200 cm, surface temperature, soil temperature at 10 cm, 40 cm, 100 cm, 200 cm, and depth of snow) across the four prespring periods. The explained variable was the Standardized Soil Moisture Index (SSMI) at 10 cm, 40 cm, 100 cm, and 200 cm soil depths.

Main Results

• The Projection Pursuit Regression (PPR) model consistently demonstrated optimal predictive performance for shallow soil layers (10 cm and 40 cm), exhibiting greater stability and a smaller fluctuation range compared to the other five models.
• For deeper soil layers (100 cm and 200 cm), the PPR model's predictive accuracy significantly decreased under high soil moisture conditions but improved markedly in dry years (e.g., 2018).
• Spatial analysis revealed that the PPR model's prediction accuracy followed the order: South Region > Central Region > North Region.
• The PPR model showed high sensitivity in identifying drought events, even under significant climate changes, but its accuracy was lower in high-humidity environments.
• Zonal predictions, particularly under drought conditions, offered advantages in drought event identification and generally demonstrated greater accuracy under normal climatic conditions compared to overall predictions.
• The Standardized Soil Moisture Index (SSMI) at different depths was best fitted by specific statistical distributions: extreme value distribution for 10 cm, normal distribution for 40 cm and 200 cm, and logit distribution for 100 cm.

Contributions

• Developed a novel and systematic method for predicting spring agricultural drought in cold and arid regions by integrating the unique characteristics of soil freeze-thaw processes during the prespring period.
• Introduced and validated the Projection Pursuit Regression (PPR) model as a superior tool for spring agricultural drought prediction in cold regions, demonstrating its effectiveness over conventional machine learning models across various soil depths and spatial scales.
• Provided a comprehensive framework for dividing the prespring period into four distinct phases (prefreezing, rapid freezing, stable freezing, thawing) and systematically selecting relevant influencing factors for drought prediction.
• Conducted a detailed comparative analysis of multiple machine learning models (PPR, PCR, PLS, PLS-PPR, LSTM, RF), offering insights into their applicability and performance under varying climatic and soil moisture conditions in cold regions.
• Enhanced the understanding of spring drought dynamics in cold regions and provided a valuable reference for developing more accurate early warning systems and drought management strategies.

Funding

• National Natural Science Foundation of China (No. 52009019)

Citation

@article{Su2026Prediction,
  author = {Su, Yi and Pei, Wei and Liu, Shaoting},
  title = {Prediction of spring agricultural drought in cold and arid regions on the basis of soil freeze thaw processes},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2026.103382},
  url = {https://doi.org/10.1016/j.ejrh.2026.103382}
}