Contourf plot with transform_first=True does not handle masked data correctly

[schlunma]

Description

Using contourf with transform_first=True on masked data simply ignores the mask and plots the underlying value.

Code to reproduce

import matplotlib.pyplot as plt
import cartopy.crs as ccrs
import numpy as np

def masked_sample_data(shape=(73, 145)):
    """Return ``lons``, ``lats`` and ``data`` of some fake data."""
    nlats, nlons = shape
    lats = np.linspace(-np.pi / 2, np.pi / 2, nlats)
    lons = np.linspace(0, 2 * np.pi, nlons)
    lons, lats = np.meshgrid(lons, lats)
    wave = 0.75 * (np.sin(2 * lats) ** 8) * np.cos(4 * lons)
    mean = 0.5 * np.cos(2 * lats) * ((np.sin(2 * lats)) ** 2 + 2)
    lats = np.rad2deg(lats)
    lons = np.rad2deg(lons)
    data = wave + mean

    # mask data
    data = np.where(data > 0.5, 1e36, data)
    data = np.ma.masked_greater(data, 10)

    return lons, lats, data

x, y, z = masked_sample_data((20, 40))

# z = np.ma.filled(z, np.nan)  # with the plot always looks good

fig = plt.figure()
projection = ccrs.Robinson(central_longitude=10.0)
axes = fig.add_subplot(projection=projection)
plot = axes.contourf(x, y, z, transform=ccrs.PlateCarree(), transform_first=True)
colorbar = fig.colorbar(plot, ax=axes)

gives a plot that shows very high values ~1e36 (which actually should be masked!):

transform_first=False and/or filling the masked array with nans gives the expected result:

Full environment definition

Operating system

Cartopy version

0.23.0

conda list

env.txt

pip list

pip.txt

[greglucas]

The transform_first is creating a new x, y, and z arrays with the definition found here:

cartopy/lib/cartopy/mpl/geoaxes.py

Lines 299 to 303 in 0fbb8a9

	This handles a fast-path optimization that projects the points before
	creating any patches or lines. This means that the lines/patches will be
	calculated in projected-space, not data-space. It requires the first
	three arguments to be x, y, and z and all must be two-dimensional to use
	the fast-path option.

It looks like we are creating new arrays for z even if you've already passed one in. I wonder if this:

cartopy/lib/cartopy/mpl/geoaxes.py

Line 315 in 0fbb8a9

x, y, z = (np.array(i) for i in args[:3])

should be np.asanyarray() instead of np.array() to respect any potential masks. I'm not sure if the take_along_axis later will respect the mask or not, but hopefully this gives someone some pointers on where to look to solve this issue.