Scatter plot with fixed markersize and hue no longer possible since v2022.10 ?

[jorisvandenbossche]

With v2022.09 and before, the scatter() plot method had the following behaviour:

import xarray as xr
temp = xr.DataArray(np.random.randn(10, 10, 2), coords=[np.arange(10), np.arange(10), [2021, 2022]], dims=["lon", "lat", "year"])
prec = xr.DataArray(np.random.randn(10, 10, 2), coords=[np.arange(10), np.arange(10), [2021, 2022]], dims=["lon", "lat", "year"])
ds = xr.Dataset({"temperature": temp, "precipitation": prec})

>>> ds
<xarray.Dataset>
Dimensions:        (lon: 10, lat: 10, year: 2)
Coordinates:
  * lon            (lon) int64 0 1 2 3 4 5 6 7 8 9
  * lat            (lat) int64 0 1 2 3 4 5 6 7 8 9
  * year           (year) int64 2021 2022
Data variables:
    temperature    (lon, lat, year) float64 0.5505 -0.2942 ... 1.085 2.287
    precipitation  (lon, lat, year) float64 0.6686 0.2782 ... 1.073 -1.053

>>> ds.plot.scatter(x="temperature", y="precipitation")

So creating a simple scatter plot for all points (across dimensions) for the two variables.

With the latest xarray, we now get a different result for the same code:

So here xarray somehow decided to use the "year" dimension for markersize, and "lat" for color (hue) (I suppose starting from the last variable / dimension not used for x/y).

This is already a (breaking, but potentially intentional) change in behaviour, but I also don't seem to be able to get the previous behaviour to use a fixed color and markersize. It seems that both keywords only accept a dimension name, and not a constant value:

>>> ds.plot.scatter(x="temperature", y="precipitation", markersize=1)
...
ValueError: size must be one of None, 'lat', 'lon', 'temperature', 'year'. Received '1' instead.

>>> ds.plot.scatter(x="temperature", y="precipitation", hue="blue")
...
ValueError: hue must be one of None, 'lat', 'lon', 'temperature', 'year'. Received 'blue' instead.

# no error, but also has no effect on actual colors (same result as above without specifying c)
>>> ds.plot.scatter(x="temperature", y="precipitation", c="blue")
<matplotlib.collections.PathCollection at 0x7f55b7109990>

Also specifying eg hue=None to explicitly say "don't use a variable to control the hue" doesn't help.

---

Looking at the whatsnew notes, this might be related to #6778? (didn't check though)

And assuming the change in behaviour was intentional, is it also on purpose there is no way to not use a variable markersizer or hue?

[Illviljan]

To make a guess which dimension to use is intentional, as there has been a precedent of guessing dimensions in other plotting methods:

import numpy as np
import matplotlib.pyplot as plt

import xarray as xr

temp = xr.DataArray(np.random.randn(10, 10, 2), coords=[np.arange(10), np.arange(10), [2021, 2022]], dims=["lon", "lat", "year"])
prec = xr.DataArray(np.random.randn(10, 10, 2), coords=[np.arange(10), np.arange(10), [2021, 2022]], dims=["lon", "lat", "year"])
one = xr.DataArray(np.random.randn(10, 10, 2), coords=[np.arange(10), np.arange(10), [2021, 2022]], dims=["lon", "lat", "year"])
blue = xr.DataArray(np.random.randn(10, 10, 2), coords=[np.arange(10), np.arange(10), [2021, 2022]], dims=["lon", "lat", "year"])
ds = xr.Dataset({"temperature": temp, "precipitation": prec, 1: one, "blue": blue})

# line adds hue:
ds.temperature.sel(lat=0).plot.line(x="lon")

# contourf adds hue:
ds.temperature.sel(year=2021).plot.contourf()

And I believe that .scatter simply didn't do this before because it was a little tricky getting categorical and numerical data to work nicely together.

You can do one of these if information about dims is of no interest:

# Stack the non interesting dims:
ds.stack(stacked_dim=[...]).plot.scatter(x="temperature", y="precipitation")

# Use matplotlib:
import matplotilib.pyplot as plt

fig, ax = plt.subplots()
ax.scatter(ds.temperature, ds.precipitation)

hue=None is the default value. If hue is none, we try to guess a value:

xarray/xarray/plot/dataarray_plot.py

Line 154 in cc7e09a

default_guess: Iterable[str] = ("x", "hue", "size"),

It's tricky to choose another default that can imply "Do not plot" however. Because xarray allows all hashables as a key.

# 1 and "blue" is hashable and maybe someone wants to use that as an array name:
fig, ax = plt.subplots()
ds.plot.scatter(x="temperature", y="precipitation", hue="blue")

[jorisvandenbossche]

I understand that the choice was made to change the default behaviour (and that this made it more consistent with other plotting methods), but I do want to say that I find it a regression in usability that it is no longer possible to specify a constant value for one of the plotting characteristics (color, size, marker) in case you case you have more than 2 dimensions in your dataset but which are not suitable to be used for those plotting characteristics (without manually having to stack those dimensions).

It's tricky to choose another default that can imply "Do not plot" however. Because xarray allows all hashables as a key.

Yes, I certainly understand, and I didn't want to suggest that this would be the best way to reuse the same keyword. I was just trying the different possibilities that might have worked.

As a reference, I checked seaborn, and they do allow to specify a fixed color and size using matplotlib's keywords for this (c/color and s in case of a scatter plot).
For an example, see one of the example towards the bottom of the scatterplot reference page (https://seaborn.pydata.org/generated/seaborn.scatterplot.html), which says:

Additional keyword arguments are passed to matplotlib.axes.Axes.scatter(), allowing you to directly set the attributes of the plot that are not semantically mapped:

sns.scatterplot(data=tips, x="total_bill", y="tip", s=100, color=".2", marker="+")

This could be something that xarray could maybe also allow? It could be possible to check if those are in kwargs and in that case not guess which dimension to use for hue and markersize and not map those to c and s, but instead use the user-specified value?

[mgunyho]

It's tricky to choose another default that can imply "Do not plot" however

I have used at least one library where None was used as the "default" and False was used for "don't do anything". Perhaps that could work here?

[dcherian]

Yeah I ran in to this too. I think we should revert the auto-guessing bit.

there has been a precedent of guessing dimensions in other plotting methods:

This precedent only applies when its unambiguous. Otherwise we error.

We should only ever make unambiguous choices, and force the user to be explicit about what they want otherwise

[stijnvanhoey]

I'm not sure if I should raise a different issue, but this seems directly linked to the fixed versus auto-guessing of hue with the scatter method:

>>> ds = xr.Dataset(data_vars=dict(pressure=(("level", "date"), np.arange(60).reshape(-1, 30)),
...                                temperature=(("level", "date"), np.random.randn(60).reshape(-1, 30))
...                               ),
...                 coords=dict(date=pd.date_range(start="2022-01-01", periods=30, freq="D"),
...                             level=np.arange(2)
...                            )
...                )
>>> ds
<xarray.Dataset>
Dimensions:      (level: 2, date: 30)
Coordinates:
  * date         (date) datetime64[ns] 2022-01-01 2022-01-02 ... 2022-01-30
  * level        (level) int64 0 1
Data variables:
    pressure     (level, date) int64 0 1 2 3 4 5 6 7 ... 52 53 54 55 56 57 58 59
    temperature  (level, date) float64 1.06 -1.668 0.3954 ... -0.7142 0.9975
>>> ds.plot.scatter(x="date", y="temperature", col="level")

returns a colorbar (auto-added with temperature data), but the colors of the scatter markers are not adjusted:

As far as I understood, I expect the color of the markers would be mapped by the temperature, i.e. current default behavior by the auto-guessing? Still, I would prefer the color mapping of the markers with another variable only when explicitly defined by the user:

>>> ds.plot.scatter(x="date", y="temperature", col="level")

no hue mapping defined, no color mapping of the markers

>>> ds.plot.scatter(x="date", y="temperature", col="level", hue="temperature")

hue mapping defined, color mapping of the markers and colorbar added