Abstract:
Reliable streamflow prediction is critical for effective water resource management, particularly in
data-scarce regions where traditional hydrological models like HEC-HMS often underperform due
to structural uncertainties and limited data. This study addresses this challenge by developing and
comparing hybrid models that integrate HEC-HMS with machine learning techniques Artificial
Neural Networks (ANN) and Extreme Gradient Boosting (XGBoost) to enhance streamflow
simulation accuracy in contrasting terrains. The Manafwa (mountainous) and Sezibwa (low-lying)
catchments in Uganda served as case studies, utilizing historical hydrological data (2010–2018)
for calibration and validation. The HEC-HMS_ANN and HEC-HMS_XGBoost models were evaluated against standalone HECHMS, ANN, and XGBoost models using Nash-Sutcliffe Efficiency (NSE), Root Mean Square
Error (RMSE), and coefficient of determination (R²). Results demonstrated that hybrid models
significantly outperformed individual approaches, with HEC-HMS_XGBoost achieving the
highest accuracy (NSE: 0.9341 calibration, 0.9191 validation for Manafwa; NSE: 0.9513
calibration, 0.937 validation for Sezibwa). The XGBoost-based hybrid excelled in capturing both
peak and low flows, while the ANN hybrid showed limitations in low-flow estimation.
Topographical differences influenced performance, with the low-lying Sezibwa catchment
exhibiting marginally better metrics than the mountainous Manafwa.
This research underscores the superiority of hybrid models, particularly HEC-HMS_XGBoost, in
overcoming data scarcity and hydrological complexity. The findings provide actionable insights
for policymakers and water managers to adopt context-specific models, enhancing flood
preparedness and infrastructure resilience in diverse catchments. The study advances hydrological
modeling in data-deficient regions, aligning with global sustainable water management goals.
