Jung et al. (2026) Development of a New Generic AI Model for Spatio‐Temporal Prediction of Soil Moisture and Soil Water Isotopes

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

Identification

• Journal: Water Resources Research
• Year: 2026
• Date: 2026-09-01
• Authors: Hyekyeng Jung, Doerthe Tetzlaff, Kristina Yordanova, Chris Soulsby
• DOI: 10.1029/2025wr041903

Research Groups

Not specified in the provided text.

Short Summary

The study implements a sequential AI approach combining LSTM and Random Forest models to simulate daily soil moisture and soil water isotopes in a mixed land use catchment, demonstrating superior performance over traditional process-based models.

Objective

• To evaluate the effectiveness of a data-driven AI approach in simulating daily soil moisture and soil water isotopes ($\delta^2\text{H}$, $\delta^{18}\text{O}$) across soil profiles using parsimonious climate and vegetation predictors.

Study Configuration

• Spatial Scale: Demnitzer Millcreek catchment (66 $\text{km}^2$), NE Germany; soil depths from 0 to 100 cm across seven land uses.
• Temporal Scale: Daily simulations; training data consisted of approximately 2 years of daily soil moisture and 1 year of monthly soil water isotopes.

Methodology and Data

• Models used: A sequential AI model consisting of a Long Short-Term Memory (LSTM) network for soil moisture prediction, which then served as an input for a Random Forest (RF) model to simulate soil water isotopes.
• Data sources: Climate and vegetation predictors; observed soil moisture and soil water isotope data.

Main Results

• Soil Moisture: Achieved Kling-Gupta Efficiency (KGE) of 0.75–0.92 and Root Mean Square Error (RMSE) of 1.81%–4.50%.
• Soil Water Isotopes: Achieved KGE of 0.80–0.88, with RMSE of 4.3–5.8‰ for $\delta^2\text{H}$ and 0.6–0.9‰ for $\delta^{18}\text{O}$.
• Comparative Performance: The AI model outperformed a previously applied process-based model in the same catchment.
• Limitations: Model performance was sensitive to the study period (overestimating in dry periods and underestimating in wet periods) and relied heavily on indirect climate predictors (e.g., relative humidity).

Contributions

• Provides a first attempt at using a sequential AI framework to simulate unsaturated zone water fluxes and mixing dynamics.
• Demonstrates the potential of AI models to serve as efficient stand-alone simulations or as surrogates for computationally intensive process-based models in ecohydrological studies.

Funding

Not specified in the provided text.

Citation

@article{Jung2026Development,
  author = {Jung, Hyekyeng and Tetzlaff, Doerthe and Yordanova, Kristina and Soulsby, Chris},
  title = {Development of a New Generic AI Model for Spatio‐Temporal Prediction of Soil Moisture and Soil Water Isotopes},
  journal = {Water Resources Research},
  year = {2026},
  doi = {10.1029/2025wr041903},
  url = {https://doi.org/10.1029/2025wr041903}
}