Dandapat et al. (2025) Ensemble deep learning framework for groundwater storage forecasting under hydrological variability

Identification

• Journal: Acta Geophysica
• Year: 2025
• Date: 2025-11-24
• Authors: Asit Kumar Dandapat, Prafulla Kumar Panda, Sovan Sankalp, Özgür Kişi, Habib Kraiem, Olga D. Kucher, Aqil Tariq
• DOI: 10.1007/s11600-025-01735-x

Research Groups

• Department of Civil Engineering, Centurion University of Technology and Management, Sitapur, Odisha, India
• Department of Civil Engineering, Gandhi Institute for Education and Technology, Nausingh, Odisha, India
• Department of Civil Engineering, Lübeck University of Applied Sciences, Lübeck, Germany
• Department of Civil Engineering, Ilia State University, Tbilisi, Georgia
• School of Civil, Environmental and Architectural Engineering, Korea University, Seoul, South Korea
• Centre for Scientific Research and Entrepreneurship, Northern Border University, Arar, Saudi Arabia
• Faculty of Artificial Intelligence, People’s Friendship University of Russia Named After Patrice Lumumba, Moscow, Russia
• Department of Wildlife, Fisheries and Aquaculture, College of Forest Resource, Mississippi State University, Mississippi State, MS, USA

Short Summary

This study develops an ensemble deep learning framework to forecast groundwater storage (GWS) in the Middle Mahanadi Basin, finding that the ensemble model significantly outperforms individual deep learning models in accuracy and reliability, particularly under hydrological variability and limited data conditions.

Objective

• To estimate monthly streamflow, analyze long-term groundwater storage trends from 1986 to 2022, and predict future groundwater storage (GWS) for 2028 in the Middle Mahanadi Basin.
• To develop and evaluate an ensemble deep learning (DL) architecture (combining LSTM, BiLSTM, SLSTM, and GRU) for accurate GWS prediction, addressing limitations of single-model forecasts and regional hydrological variability in data-scarce contexts.

Study Configuration

• Spatial Scale: Middle Mahanadi Basin, Odisha, India, with a drainage area of 9421 square kilometers, utilizing data from eight gauging stations.
• Temporal Scale: Analysis of GWS trends from 1986 to 2022, with input data (rainfall, temperature) spanning 1981 to 2022, and future GWS predictions for 2028.

Methodology and Data

• Models used:
  • Ensemble Deep Learning (EDL) model, which averages predictions from:
    • Long Short-Term Memory (LSTM)
    • Bidirectional Long Short-Term Memory (BiLSTM)
    • Stacked Long Short-Term Memory (SLSTM)
    • Gated Recurrent Unit (GRU)
  • Water budget approach for GWS estimation: ΔS = P - ET - Qs.
  • Hargreaves–Samani (HS) method for Potential Evapotranspiration (PET) estimation.
  • SCS-CN method for surface runoff estimation.
• Data sources:
  • Daily rainfall data (1981–2022) from the India Meteorological Department (IMD), New Delhi.
  • Daily maximum and minimum temperature data (1981–2022) from IMD.
  • Land use, land cover, hydrological soil condition, and antecedent soil moisture condition data for SCS-CN.
  • Data processing and analysis performed using ArcGIS and R Studio.
  • Min–max scaling for feature normalization and 80:20 training-to-test split for model training and evaluation.

Main Results

• The Ensemble DL model consistently outperformed individual LSTM, BiLSTM, SLSTM, and GRU models across all gauging stations and seasons, demonstrating the lowest Root Mean Square Error (RMSE) and highest correlation coefficient (R).
• Average GWS levels were lowest in the pre-monsoon season (304.59 mm/year), slightly increased during the monsoon (338.30 mm/year), and peaked in the post-monsoon season (653.93 mm/year).
• Seasonal trends revealed increased GWS in the post-monsoon season and declining trends during pre-monsoon and monsoon seasons.
• Significant GWS declines were observed in 2004 across most stations, attributed to reduced rainfall and increased water extraction, with slight recoveries noted in 2016 and 2022.
• Sensitivity analysis showed rainfall as the dominant driver of GWS variability (Ensemble DL: r = 0.88, NSI = 0.42), followed by runoff (r = 0.80, NSI = 0.35), and PET (r = -0.65, NSI = 0.28). Soil texture had a weaker but relevant influence (r = 0.55, NSI = 0.21).
• For the post-monsoon season, the Ensemble DL model achieved an RMSE of 0.89 and an R of 0.93, demonstrating superior accuracy.

Contributions

• Introduces an advanced ensemble deep learning framework for groundwater storage forecasting, integrating LSTM, BiLSTM, SLSTM, and GRU to overcome limitations of single-model approaches in hydrology.
• Demonstrates superior prediction accuracy and robustness of the ensemble model compared to individual deep learning models, particularly in regions characterized by high hydrological variability and limited data.
• Provides a reliable and practical tool for groundwater management and planning, offering valuable insights for effective water resource allocation and regulation in drought-prone areas.
• Enhances understanding of seasonal groundwater storage dynamics and the influence of hydroclimatic and anthropogenic factors on GWS trends in the Middle Mahanadi Basin.

Funding

• Deanship of Scientific Research at Northern Border University, Arar, KSA, through project number NBU-FPEJ-2025-2484-06.

Citation

@article{Dandapat2025Ensemble,
  author = {Dandapat, Asit Kumar and Panda, Prafulla Kumar and Sankalp, Sovan and Kişi, Özgür and Kraiem, Habib and Kucher, Olga D. and Tariq, Aqil},
  title = {Ensemble deep learning framework for groundwater storage forecasting under hydrological variability},
  journal = {Acta Geophysica},
  year = {2025},
  doi = {10.1007/s11600-025-01735-x},
  url = {https://doi.org/10.1007/s11600-025-01735-x}
}