-
Notifications
You must be signed in to change notification settings - Fork 299
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
In night-time geo_color, mid-level clouds give false impression of clear-sky conditions #2947
Comments
I definitely agree that we can improve the transparency of the high cloud layer to avoid having fully or almost fully transparent clouds. Something similar to the IR image you show would maybe be better. The same comment has been raised by Ivan who is using EUMETView, where the same recipe is used. so we could try to converge on an optimal transparency together with him, since he has a lot of experience with RGBs and also quite some interaction with users. However, I'm actually not convinced that the masked low-level clouds is (only) because of the high-level clouds. I rather think that the IR3.8-IR10.5 test used to detect low-level clouds at night struggles at twilight, when we start to get solar contribution.. Perhaps we should try using the IR3.8 BT data corrected for the solar contribution, because we have something like that in Satpy, right? I'm not sure if I agree with this being a bug, but yeah certainly room for improvement. |
This is 210/230 (this is the default): We do have the File "/data/gholl/mambaforge/envs/py312/lib/python3.12/site-packages/pyspectral/near_infrared_reflectance.py", line 267, in reflectance_from_tbs
self.derive_rad39_corr(tb_therm, tbco2)
File "/data/gholl/mambaforge/envs/py312/lib/python3.12/site-packages/pyspectral/near_infrared_reflectance.py", line 159, in derive_rad39_corr
self._rad3x_correction = (bt11 - 0.25 * (bt11 - bt13)) ** 4 / bt11 ** 4
~~~~~^~~~~~
File "/data/gholl/mambaforge/envs/py312/lib/python3.12/site-packages/dask/array/core.py", line 215, in wrapper
return f(self, other)
^^^^^^^^^^^^^^
File "/data/gholl/mambaforge/envs/py312/lib/python3.12/site-packages/dask/array/core.py", line 2400, in __sub__
return elemwise(operator.sub, self, other)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/gholl/mambaforge/envs/py312/lib/python3.12/site-packages/dask/array/core.py", line 4823, in elemwise
broadcast_shapes(*shapes)
File "/data/gholl/mambaforge/envs/py312/lib/python3.12/site-packages/dask/array/core.py", line 4751, in broadcast_shapes
raise ValueError(
ValueError: operands could not be broadcast together with shapes (11136, 11136) (5568, 5568) |
Using only FDHSI to avoid #2460, applying
|
Yeah, not great.. What repeat cycle is this? Looks like we have sun-rise to the east? Because then I don't understand why we don't have any data at night? Seems like it's being masked somehow. Perhaps we need to change the |
Interesting. Definitely get's better at some point before the low-level clouds start to become grey as well. But we probably need to keep the the use of different values (e.g. |
I think the strict limit comes from the way the emissive part is computed via first getting the reflected portion and deriving the emissive part from the reflectances. The derivation of the reflectances requires a lot of masking in Pyspectral, one of which is with the SZA. |
2024-10-15 06:00Z (repeat cycle 37). The earlier one 04:20 (RC 27).
Yes, this is the tradeoff.
Yes. My experiments were limited to Central Europe. |
If the problem is (in part) due to the emissive part of the 3,8 µm band, it might also help to change the This is what we get with Considering how much better the signal is by day as by night, it would seem we could try to use the day part closer to the terminal than SZA=78°. |
For information: in our internal production, we have now changed the parameters:
Before: After: Note that this was optimised optically for a single Central European case. We have not tested it yet on other cases or regions. |
In this geo colour example from 2024-10-28, 05:00-07:00, with my adapted settings, low clouds around and between Main and Danube are clearly shown. However, the low clouds in Styria are not only invisible due to higher clouds, but also (still) shown as clear-sky. For comparison, the version on EUMETVIEW again gives the impression the low clouds are disappearing, before reappearing: |
This case from 2024-11-11 15:30 shows that with adapted parameters (the same ones as commented three weeks ago), there is a rather dark band near the terminator. Over the North Sea, we can see high clouds that are still sunlit, but the surface is already dark. This puts us in a bit of a double bind. We want to use the solar information for the high clouds, because this has better spatial and spectral information than the thermal channels and the 3.8 µm channel is hard to use due to a mixed signal. But at the surface we would probably like to show city lights, as it's already dark there. In this case, eumetview might look better: |
Describe the bug
For the night part, the
geo_color
composite uses a blending ofgeo_color_high_clouds
andgeo_color_low_clouds
. For the high clouds, brightness temperature is shown as transparency usingHighCloudCompositor
/CloudCompositor
, where 300 K are fully transparent (not a high cloud) and a latitude-dependent threshold (either 210 K or 230 K) are fully opaque / white.When mid-level clouds cover low-level clouds, those mid-level clouds are shown mostly transparent. However, they are actually opaque, and it is unknowable whether low-level clouds exist underneath the mid-level clouds. By displaying opaque clouds as semi- or even nearly fully transparent, the user may be misled to believe there are clear-sky conditions.
An example can be seen on 2024-10-15 over Central Europe. At 04:20, low clouds cover valleys in German mountain valleys:
By 06:00, advection of mid-level clouds looks just like the fog is dissipating. Note that the mid-level clouds themselves are nearly invisible in this visualisation, but the city lights underneath are clearly shown.
After sunrise, at 07:20, we can see that actually, the low-level clouds didn't go anywhere, and are partially covered by higher clouds, now clearly visible in the true colour:
For additional views or videos, see eumetview.
To Reproduce
Expected behaviour
I expect that any (image) product clearly shows the difference between clear sky conditions and uncertain conditions. The mid-level clouds are not undetected; in fact, looking at infrared imagery at 06:00, they are clearly detectable:
So this is a visualisation problem, raising the question if it really is a good idea to scale high clouds by transparency in this case.
Actual results
See above.
Environment Info:
Additional context
I realise this problem is not easy to solve, but I think it is dangerous to present the product to forecasters in this way. One can of course warn about this problem in training, but that will get lost over time (forecasters have a lot of products to keep in mind), and in particular the city lights give the impression of clear-sky conditions where there are not (users need to remember that the city lights are a static background image). I expect users might falsely conclude that fog conditions have lifted.
The text was updated successfully, but these errors were encountered: