Wei et al. (2025) A global long-term daily multilayer soil moisture dataset derived from machine learning

Identification

• Journal: Scientific Data
• Year: 2025
• Date: 2025-12-15
• Authors: Zeyang Wei, Lifei Wei, Ting Wang, Q. Richard Lu, Shuang Tian, Fei Hu Zhang, Yanfei Zhong
• DOI: 10.1038/s41597-025-06436-0

Research Groups

• Faculty of Resources and Environmental Science, Hubei University, Wuhan, China
• Hubei Key Laboratory of Regional Development and Environmental Response, Hubei University, Wuhan, China
• Hubei Spatial Planning Research Institute, Wuhan, China
• College of Geography and Environmental Science, Zhejiang Normal University, Jinhua, China
• State Key Laboratory of Information Engineering in Surveying Mapping and Remote Sensing, Wuhan University, Wuhan, China

Short Summary

This study generated a global, daily, seamless multilayer soil moisture dataset (SWSM) for 2002–2021 at 0.05° spatial resolution using an XGBoost machine learning approach, demonstrating high accuracy against in situ observations across three soil depths. The resulting dataset addresses the scarcity of continuous, high-resolution, deep soil moisture products and provides physically consistent insights into soil moisture controls.

Objective

• To generate a global, daily, seamless multilayer soil moisture dataset (SWSM) for the period 2002–2021, at a fine spatial resolution, by leveraging a machine learning approach (XGBoost) to address the scarcity of continuous, high-resolution datasets for deep soil horizons.

Study Configuration

• Spatial Scale: Global, 0.05° spatial resolution.
• Temporal Scale: Daily, 2002–2021 (20 years).

Methodology and Data

• Models used: Extreme Gradient Boosting (XGBoost)
• Data sources:
  • Training Target: International Soil Moisture Network (ISMN) in situ observations (quality-controlled, daily averaged, depth-weighted interpolated, and normalized).
  • Input Features (Environmental Factors):
    • ERA5-Land soil moisture (primary variable)
    • ERA5-Land precipitation
    • ERA5-Land surface net solar radiation
    • GLASS Land Surface Temperature (LST)
    • GLASS Leaf Area Index (LAI)
    • MODIS/Terra+Aqua Land Cover Type (LULC)
    • Global Multi-Resolution Terrain Elevation Data 2010 (GMTED2010)
    • SoilGrids soil texture (clay, silt, sand content)
    • SoilGrids depth to bedrock (DTB)

Main Results

• The SWSM dataset provides global, daily, seamless multilayer soil moisture at 0.05° spatial resolution for three depth horizons: 0–10 cm, 10–30 cm, and 30–60 cm.
• Validation against in situ observations demonstrated high accuracy: Pearson correlation coefficients (R) exceeded 0.90 (0.905 for 0–10 cm, 0.919 for 10–30 cm and 30–60 cm) and root mean square errors (RMSE) were below 0.05 m³/m³ (0.047 m³/m³ for 0–10 cm, 0.046 m³/m³ for 10–30 cm, 0.045 m³/m³ for 30–60 cm) across all depths.
• Bias remained practically zero, and unbiased RMSE (ubRMSE) was equal to RMSE at all depths.
• SWSM exhibited superior performance over other machine learning (SoMo.ml) and model-based (GLDAS, GLEAM) datasets across various depths and soil textures in terms of median BIAS, R, and RMSE.
• Feature importance analysis (SHAP) confirmed physical consistency, revealing depth-dependent patterns where surface layers were dominated by ERA5-Land soil moisture and surface indicators (LST, LAI), while deeper layers increasingly reflected soil texture and depth to bedrock.
• Ablation analysis confirmed the fundamental contribution of ERA5-Land soil moisture and static soil properties, with surface remote-sensing indicators being more important for the top layer.
• Prediction Interval Coverage Probabilities (PICP) ranged from 93.8% to 94.5% with an average interval width of approximately 0.18, indicating reliable uncertainty representation.

Contributions

• Generation of a novel global, long-term (2002–2021), daily, seamless multilayer soil moisture dataset (SWSM) at a fine spatial resolution (0.05°), covering three distinct depths (0–10 cm, 10–30 cm, 30–60 cm).
• Development of an interpretable XGBoost-based framework that effectively integrates multi-source data for robust soil moisture estimation across depths.
• Demonstration of superior performance compared to existing global soil moisture products (SoMo.ml, GLDAS, GLEAM) in terms of accuracy and consistency.
• Provision of novel insights into depth-dependent controls on soil moisture variability through feature importance analysis, validating the physical realism of the dataset.
• The dataset serves as a valuable resource for hydrological modeling, agricultural water management, and climate change studies, addressing critical gaps in existing soil moisture products.

Funding

• National Natural Science Foundation of China (42271392)
• Opening Foundation of Xi’an Key Laboratory of Territorial Spatial Information (3001023545016)
• Open Fund of the Key Laboratory of Natural Resources Monitoring and Supervision in Southern Hilly Region, Ministry of Natural Resources (NRMSSHR2022Y02, NRMSSHR2023Y03)

Citation

@article{Wei2025global,
  author = {Wei, Zeyang and Wei, Lifei and Wang, Ting and Lu, Q. Richard and Tian, Shuang and Zhang, Fei Hu and Zhong, Yanfei},
  title = {A global long-term daily multilayer soil moisture dataset derived from machine learning},
  journal = {Scientific Data},
  year = {2025},
  doi = {10.1038/s41597-025-06436-0},
  url = {https://doi.org/10.1038/s41597-025-06436-0}
}