Li et al. (2026) Validation and ensemble-based layer-wise correction of soil moisture observations from automatic stations

Identification

• Journal: Frontiers in Environmental Science
• Year: 2026
• Date: 2026-01-14
• Authors: Huirong Li, Yaochen Li, Meng Ji, Chenlu Xu, Yongming Xu, Chunmei Wang, Moru Yan, Fan Chen, Wei Zhang
• DOI: 10.3389/fenvs.2026.1731181

Research Groups

• Xilinhot National Meteorological Observatory, China Meteorological Administration, Xilinhot, Inner Mongolia, China
• Xilinhot Field Research Station for Grassland Ecological Meteorology, China Meteorological Administration, Xilinhot, Inner Mongolia, China
• Inner Mongolia Eco-And Agro- Meteorological Centre, China Meteorological Administration, Hohhot, Inner Mongolia, China
• School of Remote Sensing and Geomatics Engineering, Nanjing University of Information Science and Technology, Nanjing, China
• Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China

Short Summary

This study validated and developed an ensemble machine learning framework for layer-wise correction of soil moisture observations from three automatic stations in Inner Mongolia, China, significantly improving their accuracy, especially in deeper soil layers.

Objective

• To validate soil moisture observations from automatic meteorological stations (DZN2, CRS-2000C, 5TM) using manual measurements as reference.
• To develop an ensemble-based, layer-wise model to correct automatic soil moisture observations across different soil depths, leveraging machine learning and environmental factors.

Study Configuration

• Spatial Scale: Xilinhot National Climate Observatory, central Xilingol Grassland of Inner Mongolia, China (43.95°N, 116.12°E; elevation: 1,124 m), a typical semi-arid grassland ecosystem.
• Temporal Scale: Data collected between May 2019 and August 2025, with varying specific periods for different sensors and data types (e.g., manual measurements from 2019-05-18 to 2025-08-21, daily or sub-daily for automatic sensors).

Methodology and Data

• Models used:
  • Base machine learning models: Cubist, Random Forest, XGBoost, CatBoost.
  • Ensemble model: Generalized Additive Model (GAM) used to integrate the predictions of the base models.
  • Hyperparameter optimization: RandomizedSearchCV with five-fold cross-validation.
• Data sources:
  • Manual ground measurements: Gravimetric soil water content (converted to volumetric water content, VWC) collected at 0–10 cm, 10–20 cm, 20–30 cm, 30–40 cm, and 40–50 cm depths.
  • Automatic soil moisture stations:
    • DZN2 automatic soil moisture sensor (FDR principle): VWC at 0–10 cm, 10–20 cm, 20–30 cm, 30–40 cm, and 40–50 cm.
    • CRS-2000C regional soil moisture measurement system (Cosmic-ray neutron sensing): VWC for the 0–50 cm soil layer.
    • 5TM soil temperature and moisture monitoring system (Capacitance/frequency-domain): VWC at 0–10 cm, 10–20 cm, and 20–40 cm.
  • Meteorological data: Daily near-surface air temperature, wind speed, relative humidity, evaporation, precipitation, and atmospheric pressure from a DZZ4 automatic weather station.
  • Remote sensing data: MODIS MOD13Q1 Normalized Difference Vegetation Index (NDVI) (250 m spatial resolution, 16-day composite).
  • Auxiliary data: Soil bulk density (for gravimetric to volumetric conversion), and calibrated soil moisture from adjacent upper layers (as input for deeper layer correction).

Main Results

• Validation: All automatic soil moisture measurements exhibited substantial biases compared to manual observations, with accuracy generally decreasing with increasing soil depth. For DZN2, R2 ranged from 0.418 (0–10 cm) to 0.179 (40–50 cm). CRS-2000C showed the highest initial accuracy with R2 of 0.774 (0–50 cm). 5TM R2 ranged from 0.673 (0–10 cm) to 0.200 (20–40 cm).
• Correction: The ensemble-based layer-wise correction framework significantly improved the accuracy of all automatic soil moisture observations.
  • R2 values increased by 0.075–0.289 across all systems and layers, with the largest increase of 0.289 for DZN2 at 30–40 cm.
  • Mean Absolute Error (MAE) decreased by 0.012–0.057 m3/m3.
  • Root Mean Square Error (RMSE) decreased by 0.013–0.075 m3/m3.
  • For DZN2, R2 increased by 62%–195%, MAE reduced by 41%–70%, and RMSE reduced by 53%–65%.
  • For CRS-2000C, R2 increased to 0.849 (a 9.7% improvement), and both MAE and RMSE were reduced by approximately 51%.
  • For 5TM, R2 increased by 18.9%–123.5%, MAE reduced by 38%–75%, and RMSE reduced by 32%–76%.
  • Improvements were particularly pronounced in deeper soil layers, and the mean bias for all soil layers approached zero after correction.
• Variable Importance: The moisture content of the immediately overlying soil layer was the most significant input variable, followed by NDVI, highlighting the effectiveness of the layer-wise approach and environmental factors.

Contributions

• Provides a comprehensive validation of three distinct automatic soil moisture monitoring systems (DZN2, CRS-2000C, 5TM) against manual ground measurements, confirming significant biases, especially at deeper soil layers.
• Develops an innovative ensemble machine learning framework, integrating Cubist, Random Forest, XGBoost, and CatBoost models via a Generalized Additive Model, to enhance correction accuracy and robustness.
• Introduces an effective layer-wise correction strategy that leverages inter-layer correlations by incorporating corrected upper-layer soil moisture as a predictor for deeper layers, addressing depth-dependent biases.
• Incorporates a broad range of meteorological and vegetation factors, along with a time-weighting approach for precipitation and evaporation, to capture complex nonlinear relationships influencing soil moisture.
• Significantly improves the accuracy and reliability of automatic soil moisture observations, particularly in deeper soil layers, providing high-precision ground-based soil moisture data crucial for drought monitoring, satellite product validation, and hydrological/climate research.

Funding

• Scientific Experiment Foundation of Inner Mongolia Meteorological Bureau (nmqxkxsy202411)
• Common Application Support Platform for National Civil Space Infrastructure “13th Five-Year Plan” Land Observation Satellites (2017–000052-73-01–001735)
• Scientific and Technological Innovation Foundation of Inner Mongolia Meteorological Bureau (nmqxkjcx202421)
• Graduate and Innovation Projects of Jiangsu Province (KYCX25_1622)

Citation

@article{Li2026Validation,
  author = {Li, Huirong and Li, Yaochen and Ji, Meng and Xu, Chenlu and Xu, Yongming and Wang, Chunmei and Yan, Moru and Chen, Fan and Zhang, Wei},
  title = {Validation and ensemble-based layer-wise correction of soil moisture observations from automatic stations},
  journal = {Frontiers in Environmental Science},
  year = {2026},
  doi = {10.3389/fenvs.2026.1731181},
  url = {https://doi.org/10.3389/fenvs.2026.1731181}
}