What should be the default behaviour of .plot.imshow() with poorly defined grids?

[KBodolai]

This is an extreme edge case, but I found the following behaviour in .plot.imshow() counterintuitive.

As a problem setting - I was testing a pipeline that combines different well-tiled arrays using xr.combine_by_coords. They form a neat, contiguous grid, but I wanted to see what happened if some were missing. The two original arrays in this example (purple and yellow) are non-contiguous in the (x,y) plane, but after being combined, they are plotted contiguously, and the x and y tick labels suggest that there is data in the 10-20 range for x and y.

Code for generating the figure

import numpy as np
import xarray as xr
import matplotlib.pyplot as plt

# Create two tiles with a gap in x
x1 = np.arange(0, 10)
x2 = np.arange(20, 30)


y1 = np.arange(0, 10)
y2 = np.arange(20, 30)

data1 = np.ones((10, 10))
data2 = np.ones((10, 10)) * 2

ds1 = xr.DataArray(data1, coords={"y": y1, "x": x1}, dims=["y", "x"]).astype(np.uint8)
ds2 = xr.DataArray(data2, coords={"y": y2, "x": x2}, dims=["y", "x"]).astype(np.uint8)

# Combine by coords
combined = xr.combine_by_coords([ds1, ds2], join='outer')

# Reindex to a full x-range, filling the gap with NaN
full_x = np.arange(0, 30)  # covers both tiles and the gap
full_y = np.arange(0, 30)
combined_fixed = combined.reindex(x=full_x, y=full_y)


fig, ax = plt.subplots(1,3, figsize=(13, 3))
combined_fixed.x.plot.line('x-', ax=ax[0], alpha=1, label='x coords after reindexing to the full range')
combined.x.plot.line('o-', ax=ax[0],alpha=0.5,  label='x coords after combine_by_coords')
ax[0].set_title('x coordinates (combined)')
ax[0].legend(loc= 'upper left')
combined.plot.imshow(ax=ax[1])
ax[1].set_title('How it displays (gap hidden)')
combined_fixed.plot.imshow(ax=ax[2])
ax[2].set_title('After reindexing')

plt.tight_layout()
plt.show()

xr.combine_by_coords here is behaving like I would expect, and plotting too, since it wraps around matplotlib.pyplot and the array is strictly contiguous (as plotted). However, the lables along the axis are misleading, since there is no data in the 10-20 range for both x and y.

This is an edge case, and arguably the problem is working with the discontinuity in the x and y coords, but to me this raises the question: should the plotter warn us that plotting discontinuous data can lead to unintuitive results?

If this is the case I'd be happy to create an issue and work in a PR to address it.

[dcherian]

IMO we can't do anything here. The grid being plotted is defined by the coordinate arrays, and non-uniformly spaced coordinate arrays are quite common.

[KBodolai]

Thanks for your reply! I mostly agree, looking at it more in depth I was starting to get that feeling as well.

To be fair, the imshow docs do have a note about this:

Note

This function needs uniformly spaced coordinates to properly label the axes. Call DataArray.plot to check.

Would it make sense to check the coordinates with xarray.utils.is_uniform_spaced and raise a warning pointing to the docs and stating that the axes will be mislabeled? On one hand I'd like xarray to tell me this more explicitly, but I'm not sure if it's the right way to do this, keen to hear your thoughts on this (and if it doesn't make sense, to mark this as answered!).

---

As a bit of a tangent, would it be useful to have is_uniform_spaced as an accessor so we can do something like:

da.is_uniform_spaced()

instead of

for coord in coords:
    xr.core.utils.is_uniform_spaced(coord)
    ...

or do we not want to expore these APIs at this level?

I'm sure that there's nunaced implementation details that I may be missing here, but would love to work on it if you think it's in-scope.