Extrapolation fixes

[pulkkins]

This pull requests fixes inconsistencies and bugs related to handling of NaN values in the semi-Lagrangian extrapolation:

• Set the keyword argument allow_nonfinite_values automatically based on the input data when calling extrapolation.semilagrangian.extrapolate from the nowcasting methods. Previously this had to be set by the user. In some nowcasting methods (e.g. ANVIL) it was hard-coded to True.
• Fix a bug in extrapolation.semilagrangian.extrapolate. The bug caused NaN values advected from outside the domain to be replaced with the minimum value when interp_order>1.

[codecov[bot]]

Codecov Report

Merging #290 (2b29e84) into master (49cc4f5) will increase coverage by 0.00%. The diff coverage is 66.66%.

@@           Coverage Diff           @@
##           master     #290   +/-   ##
=======================================
  Coverage   82.60%   82.60%           
=======================================
  Files         159      159           
  Lines       12190    12196    +6     
=======================================
+ Hits        10069    10074    +5     
- Misses       2121     2122    +1     

Flag	Coverage Δ
unit_tests	82.60% <66.66%> (+<0.01%)	:arrow_up:

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

Impacted Files	Coverage Δ
pysteps/extrapolation/semilagrangian.py	70.10% <0.00%> (-0.31%)	:arrow_down:
pysteps/nowcasts/linda.py	91.72% <50.00%> (-0.14%)	:arrow_down:
pysteps/nowcasts/anvil.py	69.73% <100.00%> (+0.23%)	:arrow_up:
pysteps/nowcasts/extrapolation.py	74.19% <100.00%> (+1.77%)	:arrow_up:
pysteps/nowcasts/sprog.py	90.30% <100.00%> (ø)
pysteps/nowcasts/sseps.py	87.41% <100.00%> (+0.02%)	:arrow_up:
pysteps/nowcasts/steps.py	86.42% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update 49cc4f5...2b29e84. Read the comment docs.

[pulkkins]

Looks all good to me! Can you just confirm my impression that this PR doesn't introduce any breaking change?

I can confirm that. I also tested all the currently implemented nowcasting methods using input data with and without NaN values. This is not included in the automatic tests, which raises a question. Should we extend the test coverage (although it might significantly increase the run time of the tests)?

[dnerini]

I also tested all the currently implemented nowcasting methods using input data with and without NaN values. This is not included in the automatic tests, which raises a question. Should we extend the test coverage (although it might significantly increase the run time of the tests)?

It’s a fair point, I’d generally go for more test coverage. If we have an issue with computation time, I think we could fairly easily make them more efficient by reducing the size of the test data. Should we take this into a separate issue or would fix that in here?

[pulkkins]

Looks all good to me! Can you just confirm my impression that this PR doesn't introduce any breaking change?

I can confirm that. I also tested all the currently implemented nowcasting methods using input data with and without NaN values. This is not included in the automatic tests, which raises a question. Should we extend the test coverage (although it might significantly increase the run time of the tests)?

It’s a fair point, I’d generally go for more test coverage. If we have an issue with computation time, I think we could fairly easily make them more efficient by reducing the size of the test data. Should we take this into a separate issue or would fix that in here?

I would take that into a separate issue. It seems that there are quite many gaps in the test coverage of the nowcasting methods.