Video summary
Supath Dhital, a full-time researcher at the Surface Dynamics Modeling Lab, presented his master's research focused on developing efficient flood inundation mapping methods to mitigate casualties caused by increasingly frequent and severe floods due to land use changes and climate change. He highlighted that while eliminating floods is impossible, accurate forecasting can significantly reduce human impact. The presentation addressed the trade-offs between different modeling approaches: remote sensing lacks forecasting capability, high-fidelity numerical models like HEC-RAS provide realistic results but are computationally expensive, low-fidelity conceptual models such as the HAND method are fast but often inaccurate and prone to misleading emergency responders, and pure data-driven machine learning models struggle with generalization when applied to new basins. To overcome these limitations, Dhital proposed a hybrid surrogate model that bridges physics-based consistency with data-driven efficiency.
The core of this research involves merging low-fidelity physically based models with response surface techniques trained on high-resolution numerical outcomes using an attention-unit convolutional neural network (CNN). The study utilized the US operational flood forecasting framework as its low-fidelity baseline, which generates synthetic rating curves from digital elevation model data to estimate inundation extent. To train this hybrid system, Dhital gathered data from 28 different watersheds across the United States featuring diverse topographies and basin characteristics. This extensive dataset allowed the model to learn complex nonlinear relationships between various hydrological inputs while maintaining physical consistency with terrain features like agricultural ditches and canals that often confuse simpler models but are critical for accurate prediction in flat, rural areas.
Validation results demonstrated that this cross-model surrogate approach significantly outperformed both pure low-fidelity operational frameworks and standard data-driven machine learning models. On the validation set, the hybrid model improved the Critical Success Index by 18% compared to the current operational framework while simultaneously reducing false alarm rates by 25%. Crucially, when tested on unseen sites across twenty different basines in the US, the median score showed a 13% improvement over benchmarks. The analysis of building-level footprints revealed that the model successfully corrected severe underpredictions and overpredictions found in low-fidelity maps; for instance, during the Minnesota River flood event, it reduced false positive counts by 57 buildings despite only minor changes to overall index scores, proving its ability to accurately delineate actual flooded structures even when general metrics fluctuate slightly.
Beyond accuracy improvements, a major advantage of this hybrid approach is its exceptional computational efficiency, making it viable for real-time operational use. While high-fidelity models require approximately 48 hours to run on standard hardware, the proposed surrogate model executed in mere seconds on a CPU and up to two thousand times faster when utilizing GPU acceleration. This dramatic speedup allows the system to be integrated directly into existing operational frameworks as a post-processor that enhances low-fidelity outputs without sacrificing physical consistency or scalability across different domains. Ultimately, Dhital concluded that this transferable hybrid model offers a practical solution for emergency response by providing enhanced flood maps quickly enough to support decision-making while accurately capturing building exposures and population risks in diverse geographical settings.
Read the full video transcript
Uh good morning everyone. Thank you for
the opportunity. Uh I'm Superal uh
full-time researcher at surface dynamics
modeling lab. So this is my part of my
master's work and and recently graduated
with my masters. Um so before starting
my presentation I want to acknowledge my
advisor uh sei kohhin and my co-authors.
So we all know like plot
plot is like most devastating natural
disaster right and common and the uh
frequency and the severity is going to
increase with the change of land and um
use and the climate change but we know
we cannot like mitigate like eliminate
the flood but we can mitigate some of
the casualties that cause due to flood.
So we can what we can do is um
floodation mapping. So it is like more
richer information than just river
disars. So what it will give you the
explicit special information about like
what where water will go. Um and it is
like very efficient and like very um I
mean helpful if you get the accurate
floodation mapping in computationally
efficient time. So my entire work and
presentation is about like how you you
can get the efficient uh flood indation
mapping which is accurate and
operationally you know feasible.
So there are different type of um
different approach to generate the
flimation map. The first one is like
remote sensing which cannot be used for
the uh forecasting purpose. The other is
like numerical models which I will call
the high fidelity models throughout the
presentation. uh which is like which
will give the realistic result but take
like computational time and the you know
resources and the other other hand there
is the low fidelity model which is like
very simple theoreion based models or
the conceptual models which is like very
fast but is not accurate and even can
mislead during the emergency response.
Uh so and the last one is like the
datadriven models uh which is kind of
popular but it's the generalization
issue with the datadriven model is like
very known so you cannot transfer from
one basin to another basin if you train
with you know in one basin. So uh it's
it's a trade-off between the accuracy
computational speed and the scalability.
Um so with that problem so um people
come up with the surrogate model
approach which is like very efficient
scalable and then can have potential to
mitigate those generalization issues
that come within the datadriven models
and which can efficiently bridge the
physics into with the data to give the
efficient result. So I'll just go
through a little bit of surrogate model
what is it? Uh so it it is of two types.
The first one is the low fidelity
physics based models. So if you know
about the so uh there is a numerical
models if you do maybe some some sort of
approximation of time consuming part
maybe you you make a coarser grid or
some you know some approximation of some
part then it becomes the low fidelity
physics based model which is consistent
throughout the topography and then uh
but not like very accurate but it's it's
like efficient and the other one is the
response surface surrogate model which
is a datadriven model but trained with
the high resolution numerical model like
outcomes. So what if we merge like both
two? So if you merge those the first low
fidelity physically based model will
give you the physically consistent like
um initial estimate of the flood
annotation mapping and then response
surface has capability to map the
nonlinear relationship between the
different input as so that it will give
the you know efficient result and it has
potential to generalization across
special domains not only the temporal
domains and um it is found to be the
like outperforming the accuracy and the
generaliz
like generalizability than the pure
derdriven like machine learning model.
So the objective of this um is like to
derive the cross model based surrogate
model because researchers have been
doing the surrogate model but they are
um getting the low fidelity and high
fidelity from the same model. Maybe the
they they develop the hairrass model
with fine c like fine grid um and then
they get the high fidelity theme and
then they just do the coarser grid to
get the low fidelity but it still take
the time to you know set up the boundary
condition and all for the hairrass. But
what if we uh try to investigate the low
fidelity from the maybe some conceptual
which is conceptual model which is
efficient or which is which are
operational um and then if we found the
result is pretty good then we can
directly integrate into the operational
models as a post-processor algorithm to
to enhance the low fidelity results and
the other is um so most of the research
have been done around the single base
like very small scale. So the one of the
another objective of this study is like
try to get as much as data from like
different basins and different
topography and then train the single
model uh that have maybe better
potential to generalize across different
uh domains and the other is like
investigate the computational efficiency
of that surrogate model because at the
end we are trying to get the efficient
model and um yeah like uh investigate
the computational efficiency
So essentially we are trying to develop
a surrogate models that mimics the
physical or numerical model attribute um
with the help of the low fidelity model.
So for the low fidelity model we going
to use the US operational um plot
forecasting framework which is the hand
height above above nearest drainage
method which is simple terbased model
and then some sort of other
topographical variables and then train
the data like the deep learning model to
develop hybrid surrogate model.
So I'll uh talk about the um different
flation mapping approach. So for the low
fidelity we're going to use the uh US
operational framework which is based on
the hand framework um and it is um
generate it generates flood in different
like four steps. First is um it will do
the hydro it will use the hydrocondition
DEM and then um derive some hydraulic
properties using that hydraulic
properties it will um it will generate
the synthetic rating curve which is
which gives the like the stage and
discharge relationship and using that
stage and discharge relationship with
topography it will just give the
inundation extent and the reason we are
like I'm interested in doing this is if
you look into the uh accuracy like the
agreement matrix that is one of like CSI
over the 49 different hydraologic unit.
um it it's just like only 0.55 for the
100red-year flood map which is really
low for the emergency uh purpose and for
my benchmarking uh my target is to like
surrogate to have like efficient um you
know um like efficiently generate the
theme right so I'll going to use the
high fidelity theme from the hairrass so
I'll use the fimma hairrass models and
use the 100 500 year return period fraud
ination maps along with that um other
synthetic flood maps which are generated
using the ripple 1D framework which is
basically raster to fim which repurpose
those femas models um to the national
water model flow lines
so about the summary of the data so in
total I'll going to train my model with
28 different sides those are basically
like 28 different hawk 8 all over the
United States those spans from like
different topography to the different
basins Um here you are you are look you
are saying those polygons are the my
modeling sides and these those point
data set are the uh my actual testing
site which will be once I train my model
I'll infer those into like 20 different
sides.
So the entire modeling will be so I'll
use the low fidelity flation mapping as
one of the predictor variable within my
attention unit based surrogate model
along with other topographical and
hydological
parameters and use the train the
attention unit convolutional neural
network which is um yeah I mean for the
special modeling the CNN is the best
and for the loss function I'm going to
use the composite loss function with
binary cross entropy along with the IOU.
So basically the composite loss function
will help us to match the pixel level
matrix as well as like boundary level
matrix. So it will help us to enhance
the result. So looking into the results
on the validation set of data. So after
applying the surrogate model, it
enhances the 18%age
um agreement between the surrogate model
with the benchmark. So on the left like
in the blue you are looking into the low
fidelity versus benchmark that is the
current operational framework versus
benchmark and on the right the orange
the surrogate model versus benchmark. So
all those metrics have like are
increased and the false alarm rate is
also decreased by 25%. But this is in
the validation set of data where I
calibrate my surrogate model. But what
about the unseen sites that I showed um
few few slides back in the data summary.
So I was interested on the unseen sides.
So I infer my surrogate the developed
surrogate model into 20 different sides
across the uh US. Those are like unseen.
Those are top like um the geography is
unseen like the temporal domain is
unseen. So here you are looking into the
change in the CSI is like mostly
positive in all of the sides. So it
means like the model is working well in
almost like all sides. Um so as a result
in those 20 sites the median score of
like 13% improvement have been uh gained
and it um predict better but uh at at
the same time in some places uh so if
you are looking into this central US
where the agricultural land and like
more flat where those false rate is a
bit high as a result the precision is
low it's because of like the
agricultural dishes and canals um has
like high topography witness index and
then what like which is one of our
strongest you know predictor in our
surrogate model that influence into that
propagate into the result. Um but those
are just like the matrix the pixel level
matrix but what about the building level
analysis because this which is very
important for the operational purpose
right. So here you are looking into the
building uh footprint analysis. So each
of the different map is um referring to
the different sides and on the left you
are looking into the low fidelity which
is current operational forecasting
framework versus the benchmark and on
the right you're looking at the
surrogate model versus the benchmark and
um you can see that like all three cases
the underprediction by the low fidelity
model that is um the Noah's hand waste
framework is severely underpredicted and
that is well captured by the surrogate
model as a result like in first Arkansas
river flood the CSI has been increased
like 27 percentage and in some cases
like it's been increased like by 46 uh
percentage and if you look at the change
in true positive building counts that
means like how many buildings that are
actually is flooded is well captured by
the surrogate model. So you can see that
like the change in the true positive
number have been increased. That means
the surrogate model is not just like
improving the metrics but it is
improving in into the right place.
Right? So um here you you are looking
into the um so those are like total
flooded building counts in three
different uh floodation map. So on the
um right like green you can see green is
the benchmark. So in the benchmark um
like the surrogate the low fidelity is
like severely underpredicted. So all in
three cases you can see in the blue like
the total number of flooded building is
very low and surrogate model is try to
bridge between the low fidelity to the
high fidelity where the yeah by
improving the flood ination map. And
other interesting case here in the um
Minnesota river flood you can see uh
there's like very small improvement in
the CSI but you can see the change in
the building footprint count. So the
false positive has been decreased by 57
uh buildings which is really huge that
means like there's very small change in
the CSI index but that translate into
the high number of like accurately
delineating the actual flooded building.
This means so if you look into the this
building footprint count the low
fidelity model is overpredicting in this
case but surrogate model is breeze
between uh to the like benchmark and
then like that over prediction is
severely cut off and you can see in the
this small enlarge panels.
So this suggests us like it can be
generalized. So with those 20 cases um
we we we get like pretty good uh like
score but what about like for the
operational purpose. So um let's talk
about the speed. So here you you are
looking into the different um um like
the computational time for the different
models. So the high fidelity that is
based on headras taking 48 hours and
hand is like low fidelity which is
pretty fast. uh but our circuit model on
the CPU is like 757
u times faster than the high fidelity
model. If you set up that into the GPU
that that becomes like 2200 uh times
faster than the high fidelity model. So
basically we are getting uh closer to
the hair accuracy um at like near you
know handbased u time speed. So that
combination is what makes this
operationally integraable and that is
the whole point of the surrogate model.
Um and we actually integrate this into
the operational framework. So how this
operation framework works is it first um
generate the low fidelity flood indation
map which is generated in the hocket
what like scale and then apply the
surrogate model with bunch of other like
attribute and the train model and it
will give the enhanced fume and down the
line like with that using that enhanced
frame it will also give the you know um
just population and the building
exposure that might be helpful though
those are not like very validated like
outcomes But it it will streamline those
process and if you are interested in the
fer you can look into this code. So the
key outcome is like hybrid surrogate
model is developed and which is
transferable into the like different
special scales um and domain
and a cross model
biases like some you know add some
complexity but you know if you once you
train the model it is like very
scalable. So you can just you know plug
into the existing framework and it will
give you the maybe enhanced result
everywhere and we get like significant
speed of ratio that help us to um that
yeah um integrate into the operational
framework.
Yeah thank you.