fix: change extrapolation values for resampling outside radar domain to zero

[RubenImhoff]

Fix for issue #416.

The small fix introduced here, which sets all extrapolation component values to zero outside the radar domain (instead of to the NWP value as in the code now), gives the following results for the case indicated in #390:

Radar observations (up to 60 minutes ahead):

Blended forecast for ensemble member 0, without the smoothing mask as introduced in #379:

Blended forecast for ensemble member 0, with the smoothing mask:

See also #390 for the previous results. The first lead time is still not perfect, but we got a lot closer already. What do you think?

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 83.88%. Comparing base (917c83b) to head (56817b3).
Report is 4 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #419   +/-   ##
=======================================
  Coverage   83.88%   83.88%           
=======================================
  Files         160      160           
  Lines       12780    12780           
=======================================
  Hits        10720    10720           
  Misses       2060     2060           

Flag	Coverage Δ
unit_tests	83.88% <100.00%> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[RubenImhoff]

@ladc, what are your thoughts here?

[ladc]

I also tried to fix this issue in the commit c7b7e0c (when addressing issue #370) but in a more complicated way. I changed the resampling to allow nan values in one of the arrays, and disregard the corresponding pixels in the resampling procedure. The results are different for the control member, but I cannot see a difference for a random member. Here's the animation comparing 447260c (first, Ruben's fix applied to my control member branch) and 7ef8a32 (last, my control member branch with modifications to resample_distributions and nonparam_match_empirical_cdf to allow nans).
In figure 1, there's a reduction in intensity in the control member for timestep 1 (maybe due to the explicit addition of zeroes in the resampling?) which is not present in figure 2, and I prefer the latter. I'm going to go through my code again to check if I haven't made any errors.

[RubenImhoff]

@ladc, I agree that your implementation would even be cleaner. If you're confident about it, could you push those resampling changes to this branch?

[ladc]

Hi everyone,
Sorry for the delay and the proliferation of branches. I've pushed my code with a ton of tests and you can find the result in the corresponding pull request #428

[dnerini]

Are we closing this in favour of #428?

[ladc]

That's OK for me.