Assessing the performance of an LSTM neural network and a HAND model for long-term real-time flood forecasting

Floods affect more people every year than any
other weather-related disaster, and can have
terrible consequences (Verwey et al., 2017). As
such, there is a need to predict flood events and
assess corresponding flood risk. If an incoming
upstream discharge wave is noticed, decisions
such as evacuation or dike reinforcement have
to be made in time and with sufficient certainty
in factors such as locations of dike overflow or
dike breaches. For this purpose, a real time
flood forecasting system is desired, in which
ensemble model predictions and uncertainty
analysis of hundreds or even thousands of
model runs allow for reviewing multiple
scenarios (Chu et al., 2020).
To model flood dynamics, the most
commonly used method is two-dimensional
depth-averaged (2DH) hydrodynamic models.
However, these models generally have long
computation times of many hours or even days.
As a result, the incoming discharge wave has
travelled further downstream, leaving little time
to evacuate the areas at risk. (Teng et al., 2017).
Surrogate models have been invented to
speed up model run times via two strategies:
lower-fidelity and response-surface surrogates
(Razavi et al., 2012). Lower-fidelity models are
still physically based, but use a simplified
description of the system. Response-surface
models are data-driven and do not contain
physical descriptions of the system.
In recent years, neural networks have
become the most popular type of responsesurface model (Mosavi et al., 2018). However,
neural networks have to be trained on a large
data-set that is often gathered from a timeconsuming 2D-hydrodynamic model, meaning
that they are only valid for the specific situation
modelled. Due to the dynamic nature of riverine
systems, both from natural and human
processes, the long-term usefulness and
validity of neural networks in a real-time flood
forecasting system is therefore questionable.
Additionally, research using neural networks for
flood water depth prediction has mostly
considered relatively simple events of rivers
spilling into floodplains. Inundation due to dike
breaches remains an unexplored field of study.
Regarding lower-fidelity models, a model
that is still being applied and developed is the
Height Above Nearest Drainage (HAND) model,
which only requires a digital elevation map of
the area to calculate maximum water depths.
Additionally, it does not require a long process
of data gathering and training, making it suitable
for easy updating in case of changing conditions
in the river system. However, it cannot predict
the propagation of the flood throughout the
study area over time, which would be useful to
decide which areas should be evacuated first.
In short, the research community finds itself
at a cross-roads: continue investing in neural
networks, or further developing lower-fidelity
models. Therefore, the objective of this
research is to expand the lower fidelity HAND
model to include a time component of flood
propagation, and compare its performance to
that of a neural network.
The two models to be developed are
evaluated against a 2D-hydrodynamic model
made in HEC-RAS by Bomers (2021). 80 model
runs that simulate the flooding after a dike
breach for different discharge waves in the
IJssel branch of the Rhine in the Netherlands
are available.