Haghighi et al. (2025) Comparative assessment of hydrological and deep learning models for runoff simulation and water storage in irrigated basins

Identification

• Journal: Modeling Earth Systems and Environment
• Year: 2025
• Date: 2025-12-06
• Authors: Alireza Razeghi Haghighi, Hossein Salehi, Ashkan Banikhedmat, Saeid Gharechelou, Rasoul Mirabbasi, Quoc Bao Pham, Ali Torabi Haghighi
• DOI: 10.1007/s40808-025-02665-9

Research Groups

• University of Tehran, Tehran, Islamic Republic of Iran
• Department of Physics, Center Agriculture, Food and Environment (C3A), University of Trento, Trento, Italy
• Shahrood University of Technology, Shahrood, Islamic Republic of Iran
• Shahrekord University, Shahrekord, Islamic Republic of Iran
• Faculty of Natural Sciences, Institute of Earth Sciences, University of Silesia in Katowice, Sosnowiec, Poland
• Water, Energy and Environmental Engineering Research Unit, University of Oulu, Oulu, Finland

Short Summary

This study evaluates the performance of physically-based and deep learning models in simulating runoff and estimating terrestrial water storage (TWS) in the Hablehroud River Basin, a semi-arid watershed in northern Iran with increasing irrigation demands. The semi-distributed Bidirectional Long Short-Term Memory (BLSTM-S) model demonstrated superior accuracy in both streamflow simulation and monthly TWS estimation, highlighting the value of deep learning in human-modified hydrological systems.

Objective

• To comprehensively compare the performance of physically-based (SWAT, VIC) and deep learning (BLSTM-L, BLSTM-S) models in simulating runoff and estimating terrestrial water storage (TWS) in the heavily irrigated, semi-arid Hablehroud River Basin.
• To assess if spatial disaggregation enhances data-driven model performance and how these approaches compare to traditional physically-based models when anthropogenic modifications significantly alter natural hydrological processes.
• To introduce a comprehensive evaluation framework that extends beyond streamflow simulation to include TWS estimation through water balance calculations, providing a more holistic assessment of model capabilities in water-stressed agricultural regions.

Study Configuration

• Spatial Scale: Hablehroud River Basin, a semi-arid watershed in northern Iran. The basin was delineated using a 30 m resolution Digital Elevation Model (SRTM). Land cover maps were at 500 m resolution (MODIS), and soil data at 1:1 million scale (HWSD). The VIC model operates on grid cells, and BLSTM-S uses a semi-distributed framework.
• Temporal Scale: The study period was from 1992 to 1996, with 1991 used for warm-up, 1992–1994 for calibration, and 1995–1996 for validation. Daily meteorological and hydrometric data were used for model inputs and runoff simulation, while Terrestrial Water Storage (TWS) was estimated monthly.

Methodology and Data

• Models used:
  • Physically-based models: SWAT (Soil and Water Assessment Tool), VIC (Variable Infiltration Capacity).
  • Deep learning models: BLSTM-L (Bidirectional Long Short-Term Memory - Lumped configuration), BLSTM-S (Bidirectional Long Short-Term Memory - Semi-distributed configuration).
• Data sources:
  • Meteorological data: Daily precipitation from 10 ground stations (Ministry of Energy), daily minimum and maximum temperatures at 2 m, and wind speed at 10 m from the ERA5 reanalysis dataset.
  • Hydrometric data: Daily streamflow observations from the Benkouh hydrometric station.
  • Physiographical data: Shuttle Radar Topography Mission (SRTM) Digital Elevation Model (DEM) (30 m resolution), Harmonized World Soil Database (HWSD) (1:1 million scale), MODIS Land Cover Type Product (MCD12Q1) (500 m resolution).
  • Evapotranspiration (ET): Monthly actual ET from the GLEAM v4.2 (Global Land Evaporation Amsterdam Model) dataset.
  • Terrestrial Water Storage (TWS): Estimated monthly as the residual of the water balance equation (Precipitation - Runoff - Evapotranspiration).

Main Results

• Daily Streamflow Simulation: The BLSTM-S model consistently outperformed all other models, achieving the highest Nash-Sutcliffe Efficiency (NSE = 0.87 during calibration, 0.80 during validation) and Kling-Gupta Efficiency (KGE = 0.91 during calibration, 0.79 during validation), with the lowest Root Mean Square Error (RMSE = 2.00 mm/day during calibration, 2.28 mm/day during validation). Physically-based models (SWAT, VIC) showed higher errors and struggled to accurately simulate low-flow conditions and streamflow recovery after irrigation periods.
• Monthly Streamflow Simulation: BLSTM-S maintained its superior performance, with NSE values of 0.93 for both calibration and validation periods, KGE values of 0.96 (calibration) and 0.86 (validation), and RMSE values of 1.21 mm/month (calibration) and 1.19 mm/month (validation). Physically-based models continued to show limitations in capturing runoff dynamics during and after the irrigation season.
• Terrestrial Water Storage (TWS) Bias: Monthly TWS estimates varied considerably across models, particularly during the growing season. BLSTM-S exhibited a narrower and more consistent bias range (typically between -2 and +2 mm/month) compared to SWAT and VIC, which showed larger seasonal biases and inconsistencies. BLSTM-L showed a notable overestimation in mid-1996.
• Annual Water Balance Components: While all models generally reproduced observed annual water balance patterns, BLSTM-S and BLSTM-L demonstrated a more balanced approach to water allocation and better represented annual hydrological variations without significant disruptions compared to physically-based models.
• Anthropogenic Influence: Physically-based models showed clear limitations in representing human-induced changes, such as irrigation withdrawals, leading to underestimation of streamflow recovery and inaccurate TWS estimates during and after irrigation periods. Deep learning models, especially BLSTM-S, demonstrated greater adaptability to these human-driven shifts in the hydrological cycle.

Contributions

• Provides the first comprehensive comparison of Bidirectional Long Short-Term Memory (BLSTM) models in both lumped and semi-distributed configurations against established physically-based models (SWAT, VIC) within a heavily irrigated semi-arid basin.
• Offers novel insights into whether spatial disaggregation enhances data-driven model performance in anthropogenically modified hydrological systems.
• Introduces a comprehensive evaluation framework that extends beyond traditional streamflow simulation to include terrestrial water storage (TWS) estimation through water balance calculations, providing a more holistic assessment of model capabilities in water-stressed agricultural regions.
• Highlights the added value of deep learning, particularly semi-distributed BLSTM, in improving both runoff simulation and seasonal water storage representation for operational water management in areas significantly affected by irrigation.
• Demonstrates that monthly TWS is a more sensitive indicator of model performance in irrigation-affected areas compared to annual estimates.

Funding

Open access funding was provided by Università degli Studi di Trento within the CRUI-CARE Agreement. This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Citation

@article{Haghighi2025Comparative,
  author = {Haghighi, Alireza Razeghi and Salehi, Hossein and Banikhedmat, Ashkan and Gharechelou, Saeid and Mirabbasi, Rasoul and Pham, Quoc Bao and Haghighi, Ali Torabi},
  title = {Comparative assessment of hydrological and deep learning models for runoff simulation and water storage in irrigated basins},
  journal = {Modeling Earth Systems and Environment},
  year = {2025},
  doi = {10.1007/s40808-025-02665-9},
  url = {https://doi.org/10.1007/s40808-025-02665-9}
}