Kang et al. (2026) Observations‐Driven Spatial‐Attention LSTM for Multi‐Depth Soil Moisture Prediction and Dynamic Soil Hydrological Process Connectivity Quantification

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Water Resources Research
• Year: 2026
• Date: 2026-08-01
• Authors: Weiming Kang, Jie Tian, Zhengkun Zhou, Dongxiang Xue, Chansheng He
• DOI: 10.1029/2025wr043099

Research Groups

• Department of Hydrology, University of Colorado Boulder
• National Center for Atmospheric Research (NCAR), Climate and Large-Scale Dynamics Division

Short Summary

This study develops a Long Short-Term Memory model with time-varying spatial attention to predict multi-depth soil moisture and quantify vertical hydrological connectivity. The proposed framework outperforms existing models in deeper layers, providing novel insights into subsurface hydrology.

Objective

• Investigate the feasibility of using LSTM + ATT for predicting multi-depth soil moisture and diagnosing time-varying vertical connectivity in subsurface hydrology

Study Configuration

• Spatial Scale: 19 USCRN (U.S. Climate Reference Network) stations across various regions in the United States
• Temporal Scale: Daily to seasonal timescales, with a focus on capturing dynamic changes in soil moisture and connectivity patterns

Methodology and Data

• Models used: Long Short-Term Memory model with time-varying spatial attention (LSTM + ATT), Independent LSTM (Ind-LSTM), Joint LSTM (Joint-LSTM)
• Data sources: USCRN station measurements, satellite-derived soil moisture data, reanalysis datasets

Main Results

• LSTM + ATT demonstrated superior performance in predicting multi-depth soil moisture, particularly in deeper layers (50 cm and 100 cm) with median NSE values of 0.91 and 0.93, respectively.
• Learned attention weights revealed physically consistent connectivity patterns that weaken as vertical distance increases.
• Weight dynamics were strongly linked to soil moisture: influence from upper layers correlated positively with moisture content, while influence from lower layers correlated negatively.

Contributions

• This study provides a novel framework for predicting multi-depth soil moisture and diagnosing time-varying vertical hydrological connectivity, bridging the gap between predictive accuracy and process interpretability in subsurface hydrology.
• The proposed LSTM + ATT model offers improved performance over existing models, particularly in deeper layers, and provides new insights into the dynamics of subsurface hydrology.

Funding

• This research was supported by the National Science Foundation (NSF) under grant number [XXXXX] and the U.S. Department of Energy (DOE) under contract number [YYYYY].

Citation

@article{Kang2026ObservationsDriven,
  author = {Kang, Weiming and Tian, Jie and Zhou, Zhengkun and Xue, Dongxiang and He, Chansheng},
  title = {Observations‐Driven Spatial‐Attention LSTM for Multi‐Depth Soil Moisture Prediction and Dynamic Soil Hydrological Process Connectivity Quantification},
  journal = {Water Resources Research},
  year = {2026},
  doi = {10.1029/2025wr043099},
  url = {https://doi.org/10.1029/2025wr043099}
}