Chen et al. (2025) A deep learning-based method for deep soil salinity prediction: considering the driving mechanisms of salinity profiles

Identification

• Journal: Geoderma
• Year: 2025
• Date: 2025-11-24
• Authors: Huifang Chen, Jingwei Wu, Chi Xu
• DOI: 10.1016/j.geoderma.2025.117615

Research Groups

• State Key Laboratory of Water Resources Engineering and Management, Wuhan University, Wuhan, Hubei, PR China
• Changjiang Institute of Survey, Planning, Design and Research, Wuhan, Hubei, PR China

Short Summary

This study investigated the transfer relationships and driving mechanisms of deep soil salinity using Hydrus-1D simulations and developed a Fully Connected Neural Network (FCNN) model to predict deep soil salinity from surface data and environmental factors, achieving R2 values from 0.44 to 0.79.

Objective

• To investigate the transfer relationships and driving mechanisms between surface salinity dynamics and deep soil salinity profiles, and to develop a method for estimating salinity at multiple depths from readily available surface salinity information.

Study Configuration

• Spatial Scale: Yichang Irrigation District of the Hetao Irrigation District, Inner Mongolia, China (41°0′40″–41°8′2″N, 107°59′33″–108°2′5″E). Soil profile depth simulated was 0.6 meters, discretized into 61 nodes with 0.01 meter spacing. Field sampling was conducted at 22 sites.
• Temporal Scale:
  • Meteorological data: 2000–2017 (11 complete years used).
  • Groundwater data: 2013.
  • Irrigation data: 2005–2020.
  • Field soil sampling: Five times in 2013 (July 7, July 28, August 8, August 28, September 22).
  • HYDRUS-1D simulation period: Non-freezing season, April 27 to November 1 each year.

Methodology and Data

• Models used:
  • HYDRUS-1D (for simulating water and solute transport in variably saturated media).
  • Fourier Amplitude Sensitivity Test (FAST) (for global sensitivity analysis).
  • Vector Autoregressive (VAR) model (for quantifying time-lagged responses).
  • Q-type cluster analysis (for classifying soil salinity profile types).
  • Random Forest (RF) (for soil salinity profile classification).
  • Fully Connected Neural Network (FCNN) (for deep soil salinity prediction).
• Data sources:
  • Field soil data: Soil samples collected from 22 sites at 0–5 cm, 5–10 cm, 10–20 cm, 20–40 cm, and 40–60 cm depths. Measurements included soil moisture content (oven-drying method) and electrical conductivity (EC1:5, μS/cm) converted to soil salt content (SSC, g/100 g).
  • Environmental data:
    • Meteorological data: Daily precipitation, sunshine duration, wind speed, relative humidity, maximum temperature, and minimum temperature from Yonglian Experimental Station.
    • Groundwater data: Groundwater tables (monitored every 5 days) and groundwater electrical conductivity (recorded every 10 days) from 10 observation wells.
    • Irrigation data: Net water intake in the Yichang Irrigation District.
  • Soil texture data: Proportions of sand, silt, and clay from the soil texture distribution map of China.
  • Simulated data: 2000 scenario combinations generated by Monte Carlo simulation and HYDRUS-1D for training and analysis.

Main Results

• The HYDRUS-1D model simulations of soil salinity demonstrated high credibility, with R2 values ranging from 0.63 to 0.74 when compared to observed data.
• Deep soil salinity dynamics are jointly driven by multiple factors (irrigation, precipitation, evapotranspiration, groundwater table, and groundwater electrical conductivity), with significant interactive effects (Total Sensitivity (ST) values ranging from 0.8 to 0.95). The independent effect of any single factor was limited (first-order sensitivity (S1) values from 0 to 0.08).
• Significant and depth-dependent lag effects were observed for various environmental factors on soil salinity:
  • Irrigation lag: 1–3 days for shallow-to-middle layers, increasing to 2–4 days for the 40–60 cm layer.
  • Precipitation lag: 1–6 days, increasing with depth and more dispersed in deeper layers.
  • Evapotranspiration lag: 2–4 days for shallow layers, 4–6 days for middle layers.
  • Groundwater table and electrical conductivity lag: 2–8 days, generally decreasing with increasing soil depth.
• Soil salinity profiles exhibited dynamic evolution, with Surface Accumulation Profile (SAP) being dominant in early periods (April–July) and late September, while a coexistence of SAP, Bottom Accumulation Profile (BAP), Middle Accumulation Profile (MAP), and Even Distribution Profile (EDP) occurred during July–September.
• Strong positive correlations were found between adjacent soil layers (e.g., 0–5 cm and 5–10 cm: r = 0.96, adjusted R2 = 0.92; 20–40 cm and 40–60 cm: r = 0.92, adjusted R2 = 0.86), which weakened with increasing depth separation (e.g., 0–5 cm and 40–60 cm: r = 0.15, adjusted R2 = 0.02).
• The Random Forest model achieved an overall accuracy of 85% in classifying soil salinity profile types, with F1-scores ranging from 0.83 (BAP) to 0.88 (SAP, EDP).
• The FCNN model effectively predicted deep soil salinity, with R2 values ranging from 0.44 to 0.79 for the validation set (using observed data) and RMSE within 0.0964 g/100 g. Prediction accuracy decreased with increasing depth (R2 for 5–10 cm: 0.79; 10–20 cm: 0.71; 20–40 cm: 0.59; 40–60 cm: 0.44).

Contributions

• Elucidated the complex transfer dynamics and driving mechanisms between surface and deep soil salinity, including the joint influence and depth-dependent lag effects of multiple environmental factors.
• Developed a novel, mechanism-constrained deep learning framework that integrates Random Forest for profile type classification and a Fully Connected Neural Network for multi-depth soil salinity prediction.
• Demonstrated the importance of incorporating soil profile characteristics and adjacent layer salinity information as covariates to enhance the accuracy and physical consistency of deep soil salinity predictions.
• Provided a new, robust framework for dynamic monitoring and management of soil salinity across different depths at a regional scale, reducing reliance on scarce deep-layer observational data.

Funding

• Key research and development program of Inner Mongolia Autonomous Region, China (2023JBGS0003)
• National Natural Science Foundation of China (Grant No. 52379047 and 52209067)
• National Key Research and Development Program of China (Grant No. 2021YFD1900804)

Citation

@article{Chen2025deep,
  author = {Chen, Huifang and Wu, Jingwei and Xu, Chi},
  title = {A deep learning-based method for deep soil salinity prediction: considering the driving mechanisms of salinity profiles},
  journal = {Geoderma},
  year = {2025},
  doi = {10.1016/j.geoderma.2025.117615},
  url = {https://doi.org/10.1016/j.geoderma.2025.117615}
}