Yan et al. (2025) Simulation of soil moisture and drought prediction in middle reaches of the Yellow River based on machine learning

Identification

• Journal: Agricultural Water Management
• Year: 2025
• Date: 2025-12-10
• Authors: Siying Yan, Baisha Weng, Zhaoyu Dong, Denghua Yan, Qiang Fu
• DOI: 10.1016/j.agwat.2025.110068

Research Groups

• State Key Laboratory of Water Cycle and Water Security, China Institute of Water Resources and Hydropower Research, Beijing, China
• School of Water Conservancy and Civil Engineering, Northeast Agricultural University, Heilongjiang, China
• State Key Laboratory of Climate System Prediction and Risk Management, Nanjing University of Information Science & Technology, Nanjing, China
• Yinshanbeilu Grassland Eco-hydrology National Observation and Research Station, China Institute of Water Resources and Hydropower Research, Beijing, China
• Institute of Environment and Ecology, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China

Short Summary

This study integrates a Multi-Layer Perceptron (MLP) model with RegCM4 climate data to generate a high-resolution, layered daily soil moisture dataset (MLP_D) for the Middle Reaches of the Yellow River (MRYR) from 2001 to 2100. Analysis of this dataset reveals a decline in deep soil moisture historically and projects a significant increase in future drought frequency and duration under intensifying climate change scenarios.

Objective

• To generate a high-resolution (0.01° × 0.01°), long-term (2001–2100), multi-depth (0–289 cm) daily soil moisture dataset (MLP_D) for the Middle Reaches of the Yellow River (MRYR) by integrating a Multi-Layer Perceptron (MLP) model with RegCM4 climate scenario data.
• To analyze the spatio-temporal distribution patterns and dynamic changes of soil moisture in the MRYR.
• To accurately predict future drought characteristics and associated risks under various climate change scenarios (RCP2.6, RCP4.5, RCP6.0, and RCP8.5) to provide a scientific basis for water resource management and drought mitigation.

Study Configuration

• Spatial Scale: Middle Reaches of the Yellow River (MRYR), covering a basin area of 344,000 km². Soil moisture profile depths: 0–7 cm, 7–28 cm, 28–100 cm, and 100–289 cm. Spatial resolution of generated MLP_D dataset: 0.01° × 0.01°.
• Temporal Scale: Historical analysis period: 2001–2022. Future projection period: 2023–2099, divided into near term (2026–2050), mid-term (2051–2075), and long term (2076–2099). All data are at a daily temporal resolution.

Methodology and Data

• Models used:
  • Machine Learning: Multi-Layer Perceptron (MLP) model for soil moisture content simulation.
  • Climate Model: RegCM4.6 regional climate model for future climate projections.
  • Interpretability: SHAP (Shapley Additive Explanations) for quantifying feature contributions.
  • Drought Assessment: Run theory for drought event identification; Relative Soil Moisture (RSM) based on China's "Drought Severity Classification Standards" (SL424–2008).
  • Trend Analysis: Linear regression method and Mann-Kendall trend test.
• Data sources:
  • Dynamic Input Data: Precipitation (Global Land Data Assimilation System, 3 h, 0.25°), Air Temperature (UK National Center for Atmospheric Science, monthly, 0.5°), Land Surface Temperature (MODIS MOD11A2.006), Normalized Difference Vegetation Index (NDVI) (MODIS MOD13Q1.006), Evapotranspiration (ET) (MODIS MOD16A2.006).
  • Static Input Data: Latitude, Longitude, Slope, Soil type, Land use, Vegetation type, Digital Elevation Model (DEM) (National Earth System Science Data Center, 30 m), Soil particle size distribution (clay, sand, silt), Soil Organic Carbon (SOC), Soil pH (U.S. Department of Agriculture, 250 m, six standard depths), Field water holding capacity (National Glacier, Permafrost, and Desert Scientific Data Center, 30″).
  • Target/Training Data: SMAP Level IV SM data (SPL4SMAU, Version 6, 2016–2022, 9 km × 9 km) for surface soil moisture (0–7 cm). ERA5-land multi-layer soil moisture data (2001–2022, 0.1° × 0.1°, 0–7 cm, 7–28 cm, 28–100 cm, 100–289 cm) for deep soil moisture.
  • Validation Data: Soil moisture station data (106 stations in MRYR, 1992–2013, 10 cm, 20 cm, 50 cm, 70 cm, 100 cm depths). GLDASNOAH0253H soil moisture data (0.25° × 0.25°, 0–10 cm, 10–40 cm, 40–100 cm, 100–200 cm). ERA5-land reanalysis data.
  • Future Climate Projection Data: RegCM4.6 model projections (average of GFDL-ESM2M, IPSL-CM5A-LR, MIROC5) under RCP2.6, RCP4.5, RCP6.0, and RCP8.5 scenarios (2007–2099, 0.25°, daily).

Main Results

• The generated MLP_D dataset (0.01° × 0.01° spatial resolution, 0–289 cm layered, daily scale, 2001–2100) demonstrates superior overall accuracy and stability (smallest bias and RMSE) compared to GLDAS and ERA5 datasets, with 90% of validation stations showing an RMSE below 0.1 m³/m³.
• During 2001–2022, surface soil moisture (0–7 cm) in the MRYR exhibited a slight, non-significant increasing trend at a rate of 0.0002 m³/m³/year. In contrast, soil moisture in layers below 7 cm declined, with a significant decrease in the 28–100 cm layer (0.0011 m³/m³/year) and a highly significant decrease in the 100–289 cm layer (0.0016 m³/m³/year).
• SHAP value analysis revealed that precipitation (SHAP=2.27%) and land surface temperature (SHAP=2.25%) are the most influential factors for surface soil moisture. For deeper soil layers, the soil moisture of the upper layer is the primary driver, followed by static topographic factors (DEM, latitude, longitude).
• The MLP_D data accurately captured the development process of the typical 2017 summer drought event, showing high consistency between simulated drought extent and actual observations.
• Under future climate scenarios (2023–2099), both the number and duration of drought events in the MRYR are projected to increase, with the trend intensifying under higher emission scenarios. Under the RCP8.5 scenario, the average annual drought duration is projected to be 95 days, and areas experiencing a significant increase in drought duration account for 71% of the total region.
• Spatially, areas with high total drought duration are concentrated in the Sanmenxia to Xiaolangdi section, Xiaolangdi to Huayuankou mainstream section, northern Fen River, and northern Yiluo River.

Contributions

• Developed and validated an innovative machine learning framework (MLP integrated with RegCM4) to generate a high-resolution, long-term, multi-depth daily soil moisture dataset (MLP_D) for the MRYR, addressing a critical data gap for drought assessment.
• Provided comprehensive insights into the spatio-temporal dynamics of soil moisture and its driving mechanisms across different depths, including the identification of key meteorological and topographic factors using SHAP analysis.
• Offered robust projections of future drought frequency, duration, and spatial distribution under various climate change scenarios, highlighting increasing drought risks, particularly under high emission pathways.
• Established a scientific basis for developing precision agriculture strategies, optimizing water resource management, and implementing effective drought mitigation policies in semi-arid regions like the MRYR.

Funding

• National Key R&D Program of China (No. 2022YFC3080300)
• Open Fund of CPRM, NUIST (No. CPRM202507)
• Research Fund of the State Key Laboratory of Simulation and Regulation of Water Cycle in River Basin (No. SKL2024YJTS03)
• IWHR Research & Development Support Program (No. MK0145B022021)

Citation

@article{Yan2025Simulation,
  author = {Yan, Siying and Weng, Baisha and Dong, Zhaoyu and Yan, Denghua and Fu, Qiang},
  title = {Simulation of soil moisture and drought prediction in middle reaches of the Yellow River based on machine learning},
  journal = {Agricultural Water Management},
  year = {2025},
  doi = {10.1016/j.agwat.2025.110068},
  url = {https://doi.org/10.1016/j.agwat.2025.110068}
}