What question did this study set out to answer?

This research aims to integrate deep learning with hydrological modeling to assess water balance components under climate variability and land use changes.

April 28, 2026Open Access

Deep Learning Algorithm and Physically‐Based Hybrid Models for Assessing Hydrological Water Balance Components Under Climate Variability and Land Use and Land Cover Change Using Earth Observations and Remote Sensing Data: The Case of Gumara Catchment, Lake Tana Sub‐Basin, Ethiopia

Key Points

This research aims to integrate deep learning with hydrological modeling to assess water balance components under climate variability and land use changes.
Developed a hybrid model using LSTM and calibrated SWAT+ outputs for hydrological analysis.
Mapped land use/land cover using a Random Forest classifier for years 1973, 1995, and 2016.
Projected future hydrological responses under SSP 245 and SSP 585 scenarios for 2050 and 2080.
Hybrid model outperformed standalone models with significant improvements in NSE, R2, and RMSE.
Direct runoff increased significantly in both historical (1973-2016) and future (2016-2080) periods, especially under high-emission scenarios.
Groundwater recharge declined notably, and streamflow varied, decreasing under SSP 245 and increasing under SSP 585.

Abstract

ABSTRACT Understanding the dynamics of basic hydrological water balance components in the Gumara Catchment, Lake Tana Sub‐Basin, Ethiopia, is crucial for climate change‐resilient agricultural planning and regional water resource management. This research presents a novel hybrid approach that integrates deep learning with process‐based hydrological modeling to evaluate the key components of the water balance under climate variability and changes in land use/land cover (LULC). Using the Quantum GIS (QGIS) interface, the Soil and Water Assessment Tool Plus (QSWAT+), and Long Short‐Term Memory (LSTM) networks, a framework is developed that leverages calibrated SWAT+ outputs as additional input features to an LSTM model. This LSTM was trained on observed historical data to estimate direct runoff, actual evapotranspiration, groundwater recharge, and streamflow for the historical years 1973, 1995, and 2016, as well as for future projections in 2050 and 2080 under both the intermediate shared socioeconomic pathway (SSP 245) and high‐emission shared socioeconomic pathway (SSP 585) scenarios. A Random Forest classifier was used to map LULC for the years 1973, 1995, and 2016. The results revealed that the hybrid model significantly outperformed the standalone LSTM and SWAT+ models over both the calibration and validation periods, as measured by Nash–Sutcliffe Efficiency (NSE), Coefficient of Determination ( R 2 ), and Root Mean‐Square Error (RMSE) performance evaluation metrics. The analysis indicated a significant increase in direct runoff in both the historical period (1973–2016) and the future period (2016–2080), particularly under high‐emission scenarios. Actual evapotranspiration showed a slight increase over the study period, while groundwater recharge declined markedly. Streamflow forecasting varied by scenario, with decreases under an intermediate pathway (SSP245) and increases under a high‐emission scenario (SSP585). These findings underscore the improved predictive ability of hybrid modeling and highlight the Gumara Catchment vulnerability to ongoing environmental change. The approach offers a sound framework to enhance water balance predictions and provides meaningful insights for sustainable water resource planning in climate‐sensitive regions like the Gumara Catchment.

Read Full Paperexternally

KI fragen

Bookmark

View Full Paper