Yin et al. (2026) Spatiotemporal prediction and attribution of groundwater storage anomaly using enhanced hybrid deep learning modeling with uncertainty quantification

Identification

• Journal: Journal of Environmental Management
• Year: 2026
• Date: 2026-01-30
• Authors: Jina Yin, Xinyao Hu, Tongchao Nan, C L Lu
• DOI: 10.1016/j.jenvman.2026.128766

Research Groups

• State Key Laboratory of Water Disaster Prevention, Hohai University, Nanjing, China
• Yangtze Institute for Conservation and Development, Hohai University, Nanjing, China
• Department of Hydrology and Water Resources, Hohai University, Nanjing, China

Short Summary

This study develops advanced hybrid deep learning models (CNN-Attention-LSTM and Transformer-LSTM) for spatiotemporal prediction, attribution, and uncertainty quantification of groundwater storage anomaly (GWSA). Applied to the Yangtze River basin, the models achieve high accuracy (R² > 0.90), attribute GWSA primarily to meteorological factors (80.66% in the middle and lower basin), and provide reliable probabilistic predictions.

Objective

• To develop and evaluate advanced hybrid deep learning models that integrate groundwater storage anomaly (GWSA) prediction, attribution analysis, and uncertainty quantification within a cohesive framework to achieve transparent and trustworthy predictions in large-scale regions.

Study Configuration

• Spatial Scale: Yangtze River basin, China.
• Temporal Scale: Prediction of groundwater storage anomaly (GWSA), focusing on capturing spatiotemporal heterogeneity and dependencies.

Methodology and Data

• Models used:
  • CNN-Attention-LSTM (CAL)
  • Transformer-LSTM (TL)
  • SHapley Additive exPlanations (SHAP) for attribution analysis
  • Stein Variational Gradient Descent (SVGD) for uncertainty quantification
• Data sources: Data-driven approach utilizing "feature data" characterized by strong spatial and temporal heterogeneity and high dimensionality, likely including meteorological and hydrological variables, for large-scale groundwater prediction.

Main Results

• Both CAL and TL hybrid deep learning models accurately predict GWSA distributions with spatiotemporal heterogeneity across the Yangtze River basin, achieving an average R² above 0.90.
• The CAL model demonstrates superior performance in localized GWSA prediction compared to TL.
• Meteorological factors are identified as the predominant contributors to GWSA, accounting for 80.66 % of the contribution in the middle and lower Yangtze River basin.
• Interactions among features and their synergistic impact on GWSA are sensitive to feature ranges.
• The incorporation of Stein Variational Gradient Descent (SVGD) significantly enhances predictive reliability, with major observations falling within the 95 % confidence intervals of the probabilistic predictions.

Contributions

• Proposes and validates advanced hybrid deep learning models (CAL and TL) that effectively capture complex spatial patterns and temporal dependencies for large-scale GWSA prediction.
• Integrates groundwater prediction, attribution analysis (using SHAP), and uncertainty quantification (using SVGD) into a single, cohesive framework, enhancing transparency and trustworthiness of predictions.
• Provides a robust methodology for interpreting intrinsic contributors to GWSA and quantifying predictive uncertainty, which is crucial for risk-informed decision-making.
• Demonstrates broad applicability and scalability of the method for other environmental tasks by allowing flexible updating of feature data.

Funding

Not explicitly mentioned in the provided paper text.

Citation

@article{Yin2026Spatiotemporal,
  author = {Yin, Jina and Hu, Xinyao and Nan, Tongchao and Lu, C L},
  title = {Spatiotemporal prediction and attribution of groundwater storage anomaly using enhanced hybrid deep learning modeling with uncertainty quantification},
  journal = {Journal of Environmental Management},
  year = {2026},
  doi = {10.1016/j.jenvman.2026.128766},
  url = {https://doi.org/10.1016/j.jenvman.2026.128766}
}