Tian et al. (2026) Deriving groundwater storage anomalies based on GRACE data and drought prediction using deep learning

Identification

• Journal: PeerJ Computer Science
• Year: 2026
• Date: 2026-01-07
• Authors: Yunna Tian, Langlang Hao, Qi Zhang, Hui Yuan, Yonghua Zhu
• DOI: 10.7717/peerj-cs.3459

Research Groups

• College of Mathematics and Computer Science, Yan’an University, China.
• College of Civil Engineering and Architecture, Yan’an University, China.

Short Summary

This study analyzes groundwater storage anomalies (GWSA) in Shaanxi Province from 2002 to 2021 using GRACE satellite and GLDAS data to establish a Standardized Groundwater Index (SGI). The research demonstrates that deep learning models, particularly the CNN-LSTM architecture, can predict groundwater drought indices with an average accuracy exceeding 84%.

Objective

• To quantify the spatial and temporal characteristics of groundwater drought in Shaanxi Province and evaluate the effectiveness of four deep learning architectures (LSTM, CNN-LSTM, BiLSTM, and LSTM-Attention) in predicting the Standardized Groundwater Index (SGI).

Study Configuration

• Spatial Scale: Shaanxi Province, China, divided into three regions (Northern Shaanxi, Guanzhong, and Southern Shaanxi) using 576 grid points at a $0.25^\circ \times 0.25^\circ$ resolution.
• Temporal Scale: April 2002 to April 2021 (Monthly resolution).

Methodology and Data

• Models used: Long Short-Term Memory (LSTM), Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM), Bidirectional Long Short-Term Memory (BiLSTM), and Long Short-Term Memory with Attention Mechanism (LSTM-Attention).
• Data sources: Gravity Recovery and Climate Experiment (GRACE) CSR Mascon data for terrestrial water storage anomalies (TWSA) and Global Land Data Assimilation System (GLDAS) Noah model for soil moisture, snow water equivalent, and canopy water.
• Statistical Analysis: Pearson Type III distribution (selected via Anderson-Darling test) was used to calculate the SGI; the Mann-Kendall (MK) test was used for trend detection.

Main Results

• Drought Trends: Northern Shaanxi and the Guanzhong region show a significant declining trend in groundwater storage, indicating worsening drought conditions. Southern Shaanxi exhibits higher variability with some areas showing increasing trends.
• Model Performance: The CNN-LSTM model achieved the highest predictive performance with a coefficient of determination ($R^2$) of 0.88, a Root Mean Squared Error (RMSE) of 0.32, and a Mean Absolute Error (MAE) of 0.16.
• Regional Accuracy: Predictive accuracy was highest in Northern Shaanxi (80–90%) due to more consistent drought patterns, while Guanzhong and Southern Shaanxi showed higher error fluctuations due to complex climatic events.
• Distribution Fitting: The Pearson Type III distribution was identified as the most robust function for modeling GWSA data across the diverse climatic zones of Shaanxi.

Contributions

• First study to systematically integrate the Standardized Groundwater Index (SGI) with multiple deep learning derivative frameworks for regional drought forecasting.
• Provides a high-resolution ($0.25^\circ$) assessment of groundwater depletion in a critical water-stressed region of China.
• Establishes a data-driven strategy for groundwater sustainability management that reduces dependency on complex physical hydrogeological modeling.

Funding

• Yan’an City Science and Technology Development Program [No. 203010096].
• Shaanxi Provincial Department of Science and Technology [No. 2023JCYB449].
• Yan’an Science and Technology Bureau [No. 2023LJBZ002 and No. 2024SLTSJJ-045].
• Graduate Education Innovation Program of Yan’an University [No. YCX2024050].
• National Natural Science Foundation of China [No. 61763046].

Citation

@article{Tian2026Deriving,
  author = {Tian, Yunna and Hao, Langlang and Zhang, Qi and Yuan, Hui and Zhu, Yonghua},
  title = {Deriving groundwater storage anomalies based on GRACE data and drought prediction using deep learning},
  journal = {PeerJ Computer Science},
  year = {2026},
  doi = {10.7717/peerj-cs.3459},
  url = {https://doi.org/10.7717/peerj-cs.3459}
}

Generated by BiblioAssistant using gemini-3-flash-preview (Google API)