Zh et al. (2026) Interpretable Groundwater-Level Prediction in an Arid Inland Basin by Integrating Dempster–Shafer Feature Screening with a Stacking Ensemble

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

Identification

• Journal: Water
• Year: 2026
• Date: 2026-07-24
• Authors: Cheng Zh, Jianhong Feng, Baohe Zhang, Liheng Wang, Yanhui Dong
• DOI: 10.3390/w18151798

Research Groups

Not specified

Short Summary

The study develops a one-day-ahead groundwater-level prediction framework for the Zhangye Basin using a stacking ensemble of deep learning models and D-S evidence theory for optimized feature selection.

Objective

• To improve the accuracy of daily groundwater-level forecasting in arid inland basins by integrating multi-source proxy variables and ensemble learning.

Study Configuration

• Spatial Scale: Zhangye Basin (focused on 10 representative wells across Zhangye and Gaotai groups).
• Temporal Scale: 2018–2025 (Daily resolution; 60-day input window).

Methodology and Data

• Models used: Dempster–Shafer (D-S) evidence theory (feature screening), Long Short-Term Memory (LSTM), Temporal Convolutional Network (TCN), Transformer (first-level sequence models), Extreme Gradient Boosting (XGBoost, second-level stacking learner), and SHapley Additive exPlanations (SHAP).
• Data sources: Groundwater level observations, meteorological data (air temperature, vapor pressure deficit), hydrological data (runoff), and proxy variables for groundwater pumping (GPV), irrigation water-demand (IWD), surface-water supply (SWS), canal-diversion (CDV), and canal irrigation supply–demand coupling intensity (CISDCI).

Main Results

• The Stacking ensemble achieved the lowest RMSE in 6 out of 10 tested wells.
• Performance metrics for the Zhangye group: RMSE = 0.1596 m, MAE = 0.0772 m, and NSE = 0.9326.
• Performance metrics for the Gaotai group: RMSE = 0.0185 m, MAE = 0.0133 m, and NSE = 0.9177.
• SHAP analysis indicated that historical groundwater levels were the dominant contributors to the model, accounting for 64.17% of the contribution in the Zhangye group and 43.96% in the Gaotai group.

Contributions

• Introduces a multi-criteria feature selection approach using D-S evidence theory to identify high-relevance process variables in arid basins.
• Demonstrates the efficacy of a stacking architecture (LSTM, TCN, Transformer $\rightarrow$ XGBoost) for short-term groundwater level prediction.
• Highlights the distinction between model dependence (on historical data) and direct hydrological causality through SHAP interpretation.

Funding

Not specified

Citation

@article{Zh2026Interpretable,
  author = {Zh, Cheng and Feng, Jianhong and Zhang, Baohe and Wang, Liheng and Dong, Yanhui},
  title = {Interpretable Groundwater-Level Prediction in an Arid Inland Basin by Integrating Dempster–Shafer Feature Screening with a Stacking Ensemble},
  journal = {Water},
  year = {2026},
  doi = {10.3390/w18151798},
  url = {https://doi.org/10.3390/w18151798}
}