Add Cell stats operator

[jfrost-mo]

What problem does your feature request solve?

Describe the solution you'd like

#2190

Describe alternatives you've considered

[A-Gainford]

After testing this branch with Bhopal radar and model data, there are a few design choices in other parts of the code that get in the way of seamless model-obs comparisons.

collapse.collapse model assumption

It is likely that cell-stats outputs will be aggregated over a full case study rather than viewed at individual snapshots. The collapse.collapse operator tries to help comparisons with multiple datasets by extracting a common forecast period/reference time from each input before collapse. Obs datasets, however, may not necessarily have these coordinates, since time is the only dimension needed to capture temporal points.

There are a few ways this could be addressed:

• Require obs datasets to include (artificial) forecast_reference_time and forecast_period coordinates. This seems undesirable, however, since it would only be needed for this one operator and doesn't meaningfully describe anything new about the data that is not already captured by the 'time' coordinate.
• Change the time extraction to use the time coordinate rather than expecting all input data to have forecast coordinates
• Make the extraction across time optional functionality of this operator.

My preference would be the latter, since I am usually already confident that the input data is valid at common times (or, at the very least, will have used an filter operator at a previous stage to enforce this).

Differences in time points separating model and obs plotting in _find_matched_slices

Unless a model has been set to output at the same timesteps as radar, it is likely that there will be a slight offset in validity times (typically models output at 5 minute intervals while radar outputs at 6 minute intervals). For evaluation across a case study, this isn't too problematic. For most of the code, this also does not present a problem.

However, plot_histogram_series separates model and obs onto different plots, as controlled by the _find_matched_slices function. This groups each cube by those with the same time points. After the collapse operator, though, the obs and models probably will not have the same point and will therefore be separated.

This seems like intended functionality, and makes sense when comparing models to other models. Could this also be an optional functionality of the code, perhaps controlled by a toggle in the operator that defaults to True to preserve existing functionality?

As a general note, I suspect there may be more of these sorts of changes that would be needed in other parts of the code. I am also aware of the nimrod branch which may have addressed some of these issues already.

[A-Gainford]

Comment lists further developments for this operator:

Dynamic vs static bins

The RMED toolbox uses a set of static bins for plotting cell stats, which works well for sub-km models but needs some adjusting for convective-scale models. Add an option that can switch between calculating bins dynamically based on input data (which np.histogram will do just fine without needing a bins to be explicitly defined) vs setting bins based on feature data type.

For effective_radius plots, this may be helped by the addition of grid_spacing to each cube output from feature.cell_stats (for calculating minimum allowable radii).

[SGallagherMet]

Comment lists further developments for this operator:

Dynamic vs static bins

The RMED toolbox uses a set of static bins for plotting cell stats, which works well for sub-km models but needs some adjusting for convective-scale models.
...

It's worth noting that even in the RMED Toolbox the choice of bins can be somewhat model dependent (for certain plot/parameter combinations). In Research to Operations we use a different set of bins for testing the UK models and the non-UK models.