Find more efficient resampler

[MarionBWeinzierl]

Working on #348, there is resampling necessary between half-hourly, hourly, a daily data. As @davidorme notes here: https://github.com/ImperialCollegeLondon/pyrealm/blob/88ed9a14b18f0e4ddf3302baa88ec7164c479a86/pyrealm_build_data/LAI_in_pyrealm/fapar_limitation_example.py#L157C1-L160C52 , the standard resampler is very slow.

An overview of alternatives can be found here: http://signalsprocessed.blogspot.com/2016/08/audio-resampling-in-python.html , and https://pypi.org/project/samplerate/ seems to be a good candidate.

[davidorme]

Looks like samplerate requires a C++ library. That might be on our development trajectory anyway, but it's a significant step 😄

[MarionBWeinzierl]

It's a simple pip install, I just did it by adding it to the .toml file and doing a poetry update. However, I am not sure whether it's more than we need?

[davidorme]

Looks like it might need a compiler on Linux?

Anyway - in general, resampling in pyrealm is either going to be:

• Averaging a bunch of values to go from fast time scales to slow ones. That is being used in that build data example ) to go from half hour to daily. We also do the same thing in subdaily model to get the average conditions in the acclimation window, but IIRC we do that in numpy with some simple reshaping along axes, which is fast.
• Interpolating a bunch of values to go from slow scales to faster ones. That's what is done in the SubdailyScaler to go from daily to subdaily value and that uses scipy.interpolate.

We aren't currently doing anything fancy with filters or smoothing and the like. I think in the build data example I was using something with a clean API, rather than a load of numpy shenanigans that would probably be far faster.

[MarionBWeinzierl]

get_window_values is doing the reshaping you mean, I believe. Is it worthwhile to extract this, and other relevant functions (like get_daily_means) out of scaler.py and make them available as core functions?

[davidorme]

It probably is, yes. I think we are always likely to be working with subdaily data at regular intervals that map neatly onto whole days (SubdailyPModel insists on this). So the reshaping and sampling there ought to be a generally useful and efficient tool?

[MarionBWeinzierl]

I created issue #406 for tackling this.

[MarionBWeinzierl]

For the second case, that just using interp1d in fill_daily_to_subdaily, I assume. But that's flagged as legacy here: https://docs.scipy.org/doc/scipy/reference/generated/scipy.interpolate.interp1d.html