Eckert et al. (2026) Soil moisture as a key predictor for regional groundwater levels: a deep learning study from Brandenburg, Germany

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

Identification

• Journal: Environmental Research Water
• Year: 2026
• Date: 2026-02-05
• Authors: Marie-Christin Eckert, née Müller Annette Rudolph
• DOI: 10.1088/3033-4942/ae4266

Research Groups

Not specified in the provided text.

Short Summary

This study developed the first regional deep learning model (1D-CNN-LSTM ensemble) for groundwater level forecasting in Brandenburg, Germany, achieving strong performance (R² = 0.72, NSE = 0.59, RMSE = 0.11) by explicitly integrating soil moisture as a key predictor, which significantly improved accuracy, especially during drought periods.

Objective

• To develop and evaluate the first regional deep learning model for groundwater levels in Brandenburg, Germany, focusing on improving forecasting accuracy by explicitly integrating vadose zone dynamics through soil moisture data.

Study Configuration

• Spatial Scale: Regional scale, Brandenburg, Germany, utilizing data from 217 monitoring wells.
• Temporal Scale: Not explicitly defined for the full study period, but covers at least 2019–2023 for evaluation of drought conditions.

Methodology and Data

• Models used: One-dimensional convolutional neural network combined with a long short-term memory (1D-CNN-LSTM) ensemble.
• Data sources: Meteorological data, soil moisture (SM) from multiple depths (0–0.3 m and 0–0.9 m), landscape predictors, and groundwater level observations from 217 monitoring wells. Utilizes nationwide open-source datasets.

Main Results

• The 1D-CNN-LSTM ensemble achieved strong regional performance with R² = 0.72, NSE = 0.59, and RMSE = 0.11.
• Explicit integration of soil moisture (SM) as a proxy for vadose zone dynamics substantially improved model accuracy.
• SM reduced bias in groundwater level estimates, particularly during the drought years 2019–2023.
• SHAP analysis identified SM from shallow (0–0.3 m) and deep (0–0.9 m) layers as the most influential predictors, surpassing all climatic inputs.
• The recurrent memory of the 1D-CNN-LSTM layers effectively captured delayed and cumulative groundwater responses across heterogeneous hydrogeological conditions.
• While individual static features had limited predictive value, their combined inclusion enhanced regional generalisation.
• Some residual overestimation persisted under extreme low-groundwater level conditions.

Contributions

• Presents the first regional deep learning model (1D-CNN-LSTM ensemble) for groundwater levels in Brandenburg, Germany.
• Introduces a key innovation by explicitly integrating soil moisture as a proxy for vadose zone dynamics, leading to substantial improvements in model accuracy.
• Demonstrates the effectiveness of 1D-CNN-LSTM in capturing complex, delayed, and cumulative groundwater responses across heterogeneous hydrogeological conditions.
• Provides a methodological framework for assessing groundwater dynamics under climate stress and guiding future improvements in regional groundwater modelling, balancing model complexity with data availability using open-source datasets.

Funding

Not specified in the provided text.

Citation

@article{Eckert2026Soil,
  author = {Eckert, Marie-Christin and Rudolph, née Müller Annette},
  title = {Soil moisture as a key predictor for regional groundwater levels: a deep learning study from Brandenburg, Germany},
  journal = {Environmental Research Water},
  year = {2026},
  doi = {10.1088/3033-4942/ae4266},
  url = {https://doi.org/10.1088/3033-4942/ae4266}
}