Strange sharp increase in return periods and flood_depth rf_glofas module

[luseverin]

I am looking into floods in Sudan and wanted to plot the maximum return period of the discharge of the entire domain over time. However, I observed a very sharp "step-function-like" increase in the return periods from 0 to 10000 years that suspiciously happen at the start of the month (01-09-2024).

However, when looking at the max discharge over time it does not look strange:

This does seem to lead to an increase in the flood depth at that time:

If you look at maps of the max return periods over the domain, it seems that the 10000 RP only occur over parts of the domain:

while the discharge map still looks fine:

Next, I show the snapshots of the maps of the return period between July 31st and August 1st. We can see that the 10000 years RP emerge over part of the domain:

@peanutfun any idea of what could be going on?

Code to reproduce the bug:

from climada_petals.hazard.rf_glofas import RiverFloodInundation
from climada_petals.hazard.rf_glofas.transform_ops import download_glofas_discharge

# download glofas discharge
iso3 = "SDN"
start_date = "2024-07-10"
end_date = "2024-09-30"
dates = pd.date_range(start_date, end_date)
query_reanal = {
    "product": "historical",
    "countries": [iso3],
    "dates": dates_event, 
}

rea_discharge_ds = download_glofas_discharge(**query_reanal)

# get return periods
rf = RiverFloodInundation()
return_periods = rf.return_period(rea_discharge_ds)

# compute flood maps
flood_ds = rf.compute(rea_discharge_ds)

# plot max return periods for the region vs time
return_periods.max(dim=("latitude", "longitude")).plot()

[peanutfun]

Did we discuss this already? I do not recall 🙈

You could have a look at the Gumbel parameters that have been calibrated there. My guess is that these are areas with very low yearly precipitation. This would make the "margin" of the discharge between low and high return periods very small. You might indeed be looking at a "feature" of EV distributions: The return period will grow exponentially (?) or at least very quickly once you get to the fringes of the distribution. Finally, note that the flood depth only related to the return period, not the discharge at all (at least in my model).

[luseverin]

Thanks Lukas! No we have not discussed this already, at least not that I can recall :) Your guess is right, Sudan is indeed dry, see plot of precipitation climatology from ERA5 (https://climateknowledgeportal.worldbank.org/country/sudan/era5-historical

)
I have looked at the loc and scale parameters and they look like this:

I am not sure on how to interpret the value of the scale parameter, but I observed negative values for the location parameter, is that physical? Could it be related to the issue we are discussing?

So if I get you correctly, it is not necessarily a bug, rather the model is not able to handle well the case of possibly extreme discharge events in a dry climate, leading to very high discharge return periods? If I am interested in looking at the return period of the discharge directly, should I just pick a point or station upstream of my region of interest and get the discharge return period timeseries from there?

[peanutfun]

Just found the computational source of the return period 10000y: This is the deliberate maximum return period the model reports. It is set here:

climada_petals/climada_petals/hazard/rf_glofas/transform_ops.py

Line 455 in 622cf98

max_return_period: float = 1e4,

During the computation of return period values, CDF values of 1.0 and larger will result in the return period being set to 10,000 years. This might be related to the negative "loc" values.

climada_petals/climada_petals/hazard/rf_glofas/transform_ops.py

Line 98 in 622cf98

cdf >= 1.0, np.inf, 1.0 / np.maximum(1.0 - cdf, np.spacing(1.0))

Your first two plots in your first comment show a more-or-less monotonic increase in discharge. It makes somewhat sense that at some point the RP then just jumps to the maximum value. But I am confused about the bump around 15 July. I don't see a decrease in discharge there. It might be so small that you cannot really see it in the plot, though.

It might make sense to look at a Gumbel distribution with negative "loc" value and see how its CDF and PDF look like.

[luseverin]

Hi,
A couple more plots; the decrease in RP indeed matches a decrease in the discharge:

and a plot of gumbel pdf and cdf with loc=-50 and scale=48:

What do you think?

[luseverin]

Sorry I did previous plot for August and not July, this is the plot for July:

So yes I don't see a decrease in discharge matching the decrease in return period; so this is indeed weird..