Jing et al. (2025) Multi-task deep learning for spatiotemporal reconstruction of groundwater dynamics in the North China Plain

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-12-22
• Authors: Hao Jing, Yong Tian, Michele Lancia, Chunmiao Zheng
• DOI: 10.1016/j.jhydrol.2025.134829

Research Groups

• School of Environment, Harbin Institute of Technology
• Guangdong-Hong Kong Joint Laboratory for Soil and Groundwater Pollution Control, School of Environmental Science & Engineering, Southern University of Science and Technology
• State Key Laboratory of Soil Pollution Control and Safety, Southern University of Science and Technology
• School of the Environment and Sustainable Engineering, Eastern Institute of Technology

Short Summary

This study developed a novel Multi-Task Learning Groundwater Model (MTLGW) integrating time-series decomposition with GRU neural networks to reconstruct regional groundwater dynamics in the North China Plain, demonstrating robust performance and superior capture of anthropogenic impacts compared to existing models.

Objective

• To develop and apply a novel Multi-Task Learning Groundwater Model (MTLGW) for spatiotemporal reconstruction of regional-scale groundwater levels in the North China Plain, particularly in data-scarce regions, to support sustainable groundwater management.

Study Configuration

• Spatial Scale: North China Plain (NCP), approximately 144000 square kilometers.
• Temporal Scale: 2005–2018.

Methodology and Data

• Models used: Multi-Task Learning Groundwater Model (MTLGW), GRU neural networks, time-series decomposition.
• Data sources: 559 groundwater monitoring wells, GLDAS-CLSM datasets.

Main Results

• The MTLGW achieved an average R² of 0.63 across the North China Plain.
• The model demonstrated strong transferability, achieving a test R² of 0.48 in regions with limited observations.
• Spatial performance varied, with higher accuracy in agricultural zones (average R² = 0.50) compared to urban areas (average R² = 0.47).
• Reconstructed shallow groundwater storage changes showed a notable mitigation of groundwater storage depletion across the NCP following the commissioning of the South-to-North Water Transfer Project (SNWT).
• MTLGW effectively captured anthropogenic impacts on groundwater dynamics, outperforming GLDAS estimates.

Contributions

• Introduces MTLGW, a novel framework integrating multi-task learning with time-series decomposition for regional-scale groundwater modeling.
• Provides a scalable framework applicable to data-limited basins worldwide.
• Bridges the gaps between AI-based predictions and sustainable groundwater management.
• Demonstrates superior capability in capturing anthropogenic impacts on groundwater dynamics compared to existing models.

Funding

• Funding information is not available in the provided text.

Citation

@article{Jing2025Multitask,
  author = {Jing, Hao and Tian, Yong and Lancia, Michele and Zheng, Chunmiao},
  title = {Multi-task deep learning for spatiotemporal reconstruction of groundwater dynamics in the North China Plain},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134829},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134829}
}