Rahman et al. (2026) Future water stress in arid landscapes projected with GeoAI

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2026
• Date: 2026-01-10
• Authors: Mahfuzur Rahman, Mahmudur Rahman, Md. Ahadul Islam Patwary, Md Masudur Rahman, Mohammed Benaafi, Isam H. Aljundi
• DOI: 10.1007/s00704-025-05976-0

Research Groups

• Interdisciplinary Research Center for Membranes and Water Security, King Fahd University of Petroleum & Minerals, Dhahran, Saudi Arabia
• Department of Chemical Engineering, King Fahd University of Petroleum & Minerals, Dhahran, Saudi Arabia
• Department of Civil Engineering, International University of Business Agriculture and Technology, Dhaka, Bangladesh
• Geosciences Department, College of Petroleum Engineering & Geosciences, King Fahd University of Petroleum & Minerals, Dhahran, Saudi Arabia
• Interdisciplinary Research Center for Aviation and Space Exploration (IRC-ASE), King Fahd University of Petroleum & Minerals, Dhahran, Saudi Arabia

Short Summary

This study develops a geospatial artificial intelligence (GeoAI) framework to project future water stress in the Eastern Province of Saudi Arabia, integrating climate models, remote sensing, and deep learning. It reveals intensifying drought severity under high-emission scenarios, particularly in summer and late-century, providing a robust tool for resilience planning.

Objective

• To develop and apply a geospatial artificial intelligence (GeoAI) framework that integrates climate model evaluation, remote sensing, deep learning, and hydroclimatic assessment to project future water stress and drought risk in the Eastern Province of Saudi Arabia, thereby supporting resilience planning in arid regions.

Study Configuration

• Spatial Scale: Eastern Province of Saudi Arabia (approximately 672,000 km²), with hydroclimatic variables mapped at 10–25 km spatial scale and land-use/land-cover (LULC) data at 10-meter spatial resolution.
• Temporal Scale: Historical baseline (2015–2024 for LULC, 2024 for CMIP6 evaluation), and future projections across four-time horizons: near future (2030), mid-century (2060), far future (2080), and end-century (2100), using daily resolution for meteorological variables.

Methodology and Data

• Models used:
  • Six CMIP6 climate models (ACCESS-ESM1-5, CanESM5, CNRM-CM6-1, EC-Earth3, ACCESS-CM2, CNRM-ESM2-1).
  • Hybrid deep learning model combining Artificial Neural Networks (ANN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU) for predicting Land Surface Temperature (LST), Potential Evapotranspiration (PET), and Actual Evapotranspiration (ET).
  • Random Forest (RF) classification model for Land Use/Land Cover (LULC) projection.
• Data sources:
  • NASA NEX-GDDP-CMIP6 dataset (bias-corrected CMIP6 global climate model projections for meteorological variables).
  • ERA5-Land reanalysis data (for CMIP6 model benchmarking).
  • Sentinel-2 Dynamic World V1 LULC dataset (for baseline and future land-use conditions).
  • NASA GLDAS Catchment Land Surface Model (CLSM, Version 2.2) groundwater storage (GWS) dataset (for independent validation of Climatic Water Availability).

Main Results

• ACCESS-ESM1-5 demonstrated superior performance among the six CMIP6 models when benchmarked against ERA5-Land (R² = 0.968, RMSE = 0.353, MAE = 0.281).
• The hybrid deep learning model accurately predicted LST, PET, and ET, achieving R² values greater than 0.94 across all scenarios, seasons, and time horizons.
• SHAP analysis confirmed the physical consistency of the deep learning models, showing wind speed, air temperature, and humidity as dominant controls on ET; radiation and temperature on PET; and air temperature and long-wave radiation on LST.
• Future projections indicate a progressive and spatially heterogeneous decline in climatic water availability (CWA) and intensifying drought severity across all emission scenarios, with the steepest reductions under SSP5-8.5.
• Under the SSP5-8.5 scenario, very-high drought severity is projected to impact over 60% of the region by the end of the century, particularly during summer and autumn.
• Spatial agreement analysis confirmed strong consistency between CWA and drought classifications (Kappa up to 0.69; Pearson r > 0.9).

Contributions

• Develops a novel, high-resolution, scenario-driven GeoAI framework that fuses CMIP6 climate projections, Sentinel-2 remote sensing, and a hybrid deep learning model (ANN-LSTM-GRU) for drought risk assessment in hyper-arid environments.
• Addresses key methodological limitations in traditional drought assessment by integrating multi-indicator, process-based water stress metrics (CWA, CWD, CWSI) and providing enhanced spatial and temporal granularity.
• Enhances the interpretability of deep learning models through SHAP analysis, validating their physical consistency with established hydroclimatic processes.
• Offers a robust, scalable, and transferable tool for anticipating drought risk and supporting adaptive land and water management in water-scarce regions, bridging the science-policy divide.

Funding

No funds, grants, or other support was received.

Citation

@article{Rahman2026Future,
  author = {Rahman, Mahfuzur and Hossain, Md Anuwer and Rahman, Mahmudur and Patwary, Md. Ahadul Islam and Rahman, Md Masudur and Benaafi, Mohammed and Aljundi, Isam H.},
  title = {Future water stress in arid landscapes projected with GeoAI},
  journal = {Theoretical and Applied Climatology},
  year = {2026},
  doi = {10.1007/s00704-025-05976-0},
  url = {https://doi.org/10.1007/s00704-025-05976-0}
}