`robust` or `clim_percentile` option to get the 2nd / 98th percentile about hvplot

[maximlt]

HoloViews has the clim_percentile plot option but that only applies to plot types that inherit from ColorbarPlot, while RGBPlot inherits from LegendPlot. @ahuang11 I'd say we should first add this plot option to HoloViews? To eventually expose it in hvPlot, where I'd say robust would be a good choice given that robust=True would be already familiar to Xarray users.

[ahuang11]

I was looking at https://github.com/holoviz/holoviews/blob/6aa6c9ddb7c9308531753efae41cbd0873c975a3/holoviews/plotting/bokeh/raster.py#L134 and I'm not sure there's a way to add clim_percentile to RGBPlot without updating Bokeh too.

I tried doing:

        if self.clim_percentile:
            low, high = 0, 1
            style["color_mapper"] = LinearColorMapper(low=low, high=high)

AttributeError [Call holoviews.ipython.show_traceback() for details]
unexpected attribute 'color_mapper' to ImageRGBA, possible attributes are anchor, decorations, dh, dh_units, dilate, dw, dw_units, global_alpha, image, js_event_callbacks, js_property_callbacks, name, origin, subscribed_events, syncable, tags, x or y

I also tried doing:

        if self.clim_percentile:
            low, high = np.nanpercentile(img, (2, 98))
            img[img < low] = low
            img[img > high] = high

But it doesn't seem to work

[ahuang11]

Thoughts about having this in hvPlot?
holoviz/holoviews#6137

I don't have a strong preference, but I do want to have a way forward.

[maximlt]

I also don't have strong preference :)

[Hoxbro]

I also don't have strong preference 🙃

[philippjfr]

I'm -1 on adding this anywhere. Taking percentiles of RGB values just does not make sense to me.

[ahuang11]

If that's the case, how should users address this?

The initial goal was that I wanted to write a demo of how you can explore satellite imagery with hvplot easily, but then I encountered the issue that robust/clim_percentile was not exposed.

If robust/clim_percentile isn't implemented, its a tad tedious for the user to have to manually implement it for satellite imagery vs typing robust=True.

Would you prefer it be an operation? E.g. Clip?

[philippjfr]

Will have to look at it.

[philippjfr]

Yeah, looking at this dataset the values are not R G B values they are raw reflectance values for different spectra from a satellite. Visualizing them using an RGB element without first processing these values is simply incorrect. So the only thing I can imagine doing is providing some operation in geoviews.

[philippjfr]

https://ecampusontario.pressbooks.pub/remotesensing/chapter/chapter-5-visualization-and-manipulation-of-satellite-images/

[philippjfr]

Based on my reading of that the transformation is quite simple, you divide the red green blue values 2**8 and then cast to uint8.

[philippjfr]

Alternatively you can also apply histogram equalization.

[philippjfr]

Hmm, the division thing doesn't work.

from datashader.transfer_functions import eq_hist
hv.RGB(eq_hist(da.data)[0].transpose([1, 2, 0])).opts(width=600)

[philippjfr]

hv.RGB((da / (2**2)).clip(0, 255).astype(np.uint8).data.transpose([1, 2, 0])).opts(width=600)

[philippjfr]

Okay, I've come around I'd be okay with a two-pronged approach, we implement the robust semantics in hvPlot and then HoloViews appropriately clips the data.

[ahuang11]

Philipp found the actual _rescale_imshow_rgb robust usage in xarray here