Machine-learning projections of groundwater storage under climate, land-use, and human-driven changes in Madhesh Province, Nepal
- 1Environmental Engineering Program, Department of Civil Engineering, Institute of Engineering, Tribhuvan University
- 2Department of Civil Engineering, Institute of Engineering, Tribhuvan University
Abstract
Groundwater is the primary source of water in Madhesh Province, Nepal, yet the province lacks a reliable forecasting framework. This study develops a provincial-scale groundwater forecasting system by integrating satellite-derived groundwater storage (GWS), machine learning, and CMIP6 climate forcing. A Long Short-Term Memory (LSTM) model was trained using two decades of GRACE–GLDAS anomalies, observed climate data, and land use–land cover (LULC) information, and was forced with bias-corrected MIROC6 projections under the SSP5-8.5 scenario. Results reveal a persistent historical GWS decline of approximately 50 mm over 20 years (2.5 mm yr⁻¹), driven primarily by rainfall variability and intensified by urban expansion and surface water loss. Two LSTM setups were tested: a climate-driven model using precipitation and temperature, which achieved higher predictive accuracy but underrepresented human influences, and a multi-parameter model incorporating LULC, soil moisture, groundwater-irrigated area, and domestic demand, which captured more realistic depletion dynamics despite slightly lower statistical performance. Both the climate-driven and multi-parameter LSTM models indicate a monsoon-dependent, highly seasonal recharge with weakening peaks and continued decline through 2045. Furthermore, scenario-based demand sensitivity showed that mean GWS increased from 635.20 mm to 639.82 mm under demand − 10%, but declined to 623.91 mm under total demand + 30%, showing a mean 95% confidence interval width of 9.24 mm and a mean model spread of 7.78 mm, indicating stable projection responses across model structures. The bias-corrected Machine Learning-driven GWS and predictor dataset provides a reproducible basis for groundwater assessments and model benchmarking in data scare regions. These findings provide actionable insights for groundwater management, supporting climate-resilient water resource planning, sustainable irrigation practices, and policy development in Madhesh Province and similar data-scarce alluvial regions.
Related
- Monitoring and forecasting groundwater storage dynamics in arid Wadi systems using GRACE/GRACE-FO satellite observations and XGBoost machine learning: Wadi El-Assiuti, Egypt (2003–2030)
- Global projections of groundwater sustainability through the end of the century
- Defining and predicting groundwater drought dynamics in a temperate maritime region using Standardised Groundwater Index (SGI) and ensemble machine learning
- A National Scale Hybrid Machine Learning Model for Enhanced Groundwater Depth Estimation
- Groundwater potential mapping using conventional and machine learning approaches with multi-indicator validation of Aflaj systems in Oman
BibTeX
@article{mahto2026machinelearning,
title = {{Machine-learning projections of groundwater storage under climate, land-use, and human-driven changes in Madhesh Province, Nepal}},
author = {Anjana Kumari Mahto and Vishan Dahal and Ram Krishna Regmi and Govinda Prasad Poudel and Ramesh Karki and Prakash Chandra Ghimire and Shukra Raj Paudel},
journal = {Scientific Reports},
year = {2026},
month = sep,
doi = {10.1038/s41598-026-70415-7},
url = {https://doi.org/10.1038/s41598-026-70415-7},
note = {CC BY},
}
RIS
TY - JOUR
TI - Machine-learning projections of groundwater storage under climate, land-use, and human-driven changes in Madhesh Province, Nepal
AU - Anjana Kumari Mahto
AU - Vishan Dahal
AU - Ram Krishna Regmi
AU - Govinda Prasad Poudel
AU - Ramesh Karki
AU - Prakash Chandra Ghimire
AU - Shukra Raj Paudel
PY - 2026
DA - 2026/09/01
JO - Scientific Reports
DO - 10.1038/s41598-026-70415-7
UR - https://doi.org/10.1038/s41598-026-70415-7
AB - Groundwater is the primary source of water in Madhesh Province, Nepal, yet the province lacks a reliable forecasting framework. This study develops a provincial-scale groundwater forecasting system by integrating satellite-derived groundwater storage (GWS), machine learning, and CMIP6 climate forcing. A Long Short-Term Memory (LSTM) model was trained using two decades of GRACE–GLDAS anomalies, observed climate data, and land use–land cover (LULC) information, and was forced with bias-corrected MIROC6 projections under the SSP5-8.5 scenario. Results reveal a persistent historical GWS decline of approximately 50 mm over 20 years (2.5 mm yr⁻¹), driven primarily by rainfall variability and intensified by urban expansion and surface water loss. Two LSTM setups were tested: a climate-driven model using precipitation and temperature, which achieved higher predictive accuracy but underrepresented human influences, and a multi-parameter model incorporating LULC, soil moisture, groundwater-irrigated area, and domestic demand, which captured more realistic depletion dynamics despite slightly lower statistical performance. Both the climate-driven and multi-parameter LSTM models indicate a monsoon-dependent, highly seasonal recharge with weakening peaks and continued decline through 2045. Furthermore, scenario-based demand sensitivity showed that mean GWS increased from 635.20 mm to 639.82 mm under demand − 10%, but declined to 623.91 mm under total demand + 30%, showing a mean 95% confidence interval width of 9.24 mm and a mean model spread of 7.78 mm, indicating stable projection responses across model structures. The bias-corrected Machine Learning-driven GWS and predictor dataset provides a reproducible basis for groundwater assessments and model benchmarking in data scare regions. These findings provide actionable insights for groundwater management, supporting climate-resilient water resource planning, sustainable irrigation practices, and policy development in Madhesh Province and similar data-scarce alluvial regions.
SN - 2045-2322
ER -