ANN–HEC-RAS Integrated Modeling for Flood Inundation Mapping in the Bagmati River at Khokana, Nepal

Abstract

Floods in the Bagmati river pose serious risks to the
residents of Khokana, with their floodplain gradually being
encroached on due to urbanization. Modeling floods in areas
that lack sufficient hydrological measurements and
topographical information can be difficult. In this paper, an
Artificial Neural Network (ANN) rainfall–runoff model was
combined with a 2-dimensional (2D) hydraulic HEC-RAS
model to provide high-resolution flood inundation maps of the
Bagmati River reach in Khokana. Seven-year daily hydrometeorological
data (2019–2025) were applied to train (2019–
2023), validate (2024), and test (2025) chronologically without
using early stopping and/or model selection on any other period
except the validation year in order to avoid overfitting. A
feedforward multilayer perceptron (MLP) neural network,
trained to minimize a percentile-targeted, peak-weighted loss
function, scored a Nash-Sutcliffe Efficiency (NSE) of 0.86 and a
peak-flow percent bias (PBIAS) of 2.0% for the independent
year 2025 test. A recurrent Long Short-Term Memory (LSTM)
neural network model was further compared to the MLP model
under the same chronological approach. A 14.9 cm highresolution
Digital Terrain Model (DTM), obtained from a UAV,
was employed for capturing micro-topography of the
embankments and floodplains, and the flood extents delineated
using this model were compared against the satellite-derived
Normalized Difference Water Index (NDWI) flood extents
obtained from Sentinel-2 at the 10 m native resolution.
Furthermore, discharge sensitivity analysis helps to quantify
the propagation of residual peak flow errors from ANN to
inundation extent and depth, thereby making a clear connection
between the hydrological model performance and the
consequence of the flood mapping. The results demonstrate that
an ANN–HEC-RAS framework, rigorously validated with a
held-out test year, can support disaster risk reduction and landuse
planning in data-scarce, rapidly urbanizing floodplains
such as Khokana.