Abstract:
The Kalu River Basin in southwestern Sri
Lanka has experienced significant flooding, primarily
from monsoonal and cyclonic rainfall. Steep topography,
rapid land-use changes, and climate change exacerbate
flood risk. Notable flood events in this region occurred in
2003, 2008, 2012, 2014, 2017, 2018, and 2021, causing
substantial economic losses and social impacts.
Traditional forecasting methods have relied on physically
based hydrological and hydrodynamic models to
simulate rainfall–runoff relationships, particularly in
well-instrumented catchments. However, these
conventional models often require extensive data,
presenting challenges in data-limited environments such
as Sri Lanka. Machine learning (ML) models offer an
alternative by learning flood-generating patterns directly
from hydrological and environmental data, achieving
high forecasting accuracy with limited information.
In contrast, process-based models encounter
challenges such as inflexibility, nonlinear dynamics, and
residual biases. This paper reviews global and regional
research focusing on hybrid approaches that integrate
physical models with ML techniques and assesses their
potential application in the Kalu River Basin. The review
emphasizes the significance of combining physical and
ML models, demonstrating that hybrid methods generally
outperform standalone models in estimating peak flows
and predicting extreme events, even under data-limited
conditions. In summary, this review advocates adopting
integrated modeling systems as robust and adaptable
solutions for flood forecasting.
