Bugfix forced log scales in raster plots

[ChrisFairless]

Changes proposed in this PR:

I hope you're ready for another PR from me where the explanation is 20 times longer than the actual fix!

Overview:

When you want to plot a raster for an Exposure or an Impact, so that plot

• doesn't plot zero values, and
• applies a non-default function to the raster values before plotting (i.e. not log-scaling),

then CLIMADA unintentionally overrides your choice of function.

More precisely, when Exposures.plot_raster() has fill=False, it overrides the user-provided raster_f parameter.

In detail:

In the method, raster_f is a lambda function that transforms the value of each raster cell and is logarithmic by default (raster_f=lambda x: np.log10((np.fmax(x + 1, 1)))). This bug means you can't actually plot a raster with its raw values (i.e. with raster_f: lambda x: x).

The culprit is in lines 1063-1065:

        if not fill:
            raster = np.where(raster == 0, np.nan, raster)
            raster_f = lambda x: np.log10((np.maximum(x + 1, 1)))

The intention made sense: the default parameters with fill=True use raster_f=lambda x: np.log10((np.fmax(x + 1, 1))). This lambda uses np.fmax, meaning raster_f(np.nan) = 0 meaning that missing values are treated as zeroes and coloured accordingly.

But when fill=False the lines above kick in and override with a new raster_f=lambda x: np.log10((np.maximum(x + 1, 1))). Using np.maximum now means raster_f(np.nan) = np.nan instead, and therefore missing values (and zeroes, due to the mask applied one line earlier) are treated as missing. Which was a neat trick. But it forget that raster_f could be any function, and we don't actually to override it.

The fix:

The solution is just to apply a missing value mask if fill=False instead.

Aside: but wait, what is the method trying to do?

It took me a while to work out the intended behaviour:

• fill=True: np.nan -> 0
• fill=False: np.nan -> np.nan
• fill=True: 0 -> 0
• fill=False: 0 -> ?
  According to the docstring, this should be 0 BUT the code explicitly masks them, making them np.nan.

I assume the code is exactly what is intended and we always want to mask zero values when fill=False. This makes sense when plotting financially-valued assets.

So that's what I've implemented (fill=False: 0 -> np.nan ) and I've updated the docstring.

Finally: I think what this function actually wanted to accomplish all this time was a mpl.colors.LogNorm colourbar, like in plot.plot_from_gdf. But let's not go into that.

Tests:

Codex wrote an obscure test that took me 10 minutes to understand so I thought it wasn't worth including. A smarter approach would be to move all the masking and transforming into a separate helper function and test that, but I'm not sure we really need it, and this PR is long enough already???

PR Author Checklist

• Read the Contribution Guide
• Correct target branch selected (if unsure, select develop)
• Descriptive pull request title added
• Source branch up-to-date with target branch
• Documentation updated
• Tests updated
• Tests passing
• No new linter issues
• Changelog updated

PR Reviewer Checklist

• Read the Contribution Guide
• CLIMADA Reviewer Checklist passed
• Tests passing
• No new linter issues