Remove esri.py and datashader from core dependencies

[brendancol]

Author of Proposal: @brendancol

Reason or problem

datashader is a core dependency in setup.cfg, but the library barely uses it:

• terrain.py and datasets/__init__.py use ds.Canvas.points on an empty DataFrame just to generate coordinate arrays (there's a comment calling it hack_agg)
• utils.py wraps ds.Canvas.raster in canvas_like() and uses tf.Image in two visualization helpers (bands_to_img, color_values)
• test_viewshed.py uses ds.Canvas to build a test fixture

Datashader drags in bokeh, colorcet, panel, param, and more. That's a lot of transitive weight for users who just want raster math.

Separately, esri.py has REST-client utilities for querying ArcGIS feature services. Nothing in the codebase imports it, __init__.py doesn't re-export it, and there are no tests. It also pulls in requests as an undeclared dependency. It shouldn't be here.

Proposal

Design:

1. Delete esri.py. Nothing references it.

2. Remove datashader from install_requires in setup.cfg. Add it to [options.extras_require] examples so notebook users can still get it via pip install xarray-spatial[examples].

3. Replace the ds.Canvas coordinate hacks in terrain.py and datasets/__init__.py with np.linspace. The current code creates a Canvas, calls .points() on an empty DataFrame, and steals the coords from the result. We can build the coordinate arrays directly.

4. Swap tf.Image in bands_to_img() and color_values() for PIL.Image (Pillow), already available via matplotlib.

5. Replace datashader.colors.rgb() in color_values() with a local hex-to-RGB helper.

6. Make canvas_like() lazy-import datashader and raise a clear ImportError if missing. This function wraps cvs.raster() and there's no drop-in replacement.

7. Update test_viewshed.py to build its fixture DataArray with np.linspace coordinates.

8. Update the docstring example in zonal.py to note that datashader is optional.

Usage:

pip install xarray-spatial will no longer pull in datashader. Users who need canvas_like() or the example notebooks can install extras: pip install xarray-spatial[examples].

Value:

Lighter install and fewer version conflicts. Removes dead code.

Stakeholders and impacts

All downstream users get a lighter install. The only functions affected are canvas_like, bands_to_img, and color_values -- visualization helpers, not core spatial algorithms.

Drawbacks

• canvas_like will raise ImportError if datashader isn't installed. Users who depend on it need to add datashader to their own requirements.
• bands_to_img and color_values will return PIL.Image objects instead of datashader.transfer_functions.Image. Both are PIL subclasses in practice, so most downstream code should work fine.

Alternatives

• Lazy-import datashader everywhere with try/except. Adds complexity for little benefit when the core library doesn't need it.
• Move datashader to extras_require only. That's what this proposal does for canvas_like; the rest gets replaced outright.

Unresolved questions

None.

Additional notes

Removing esri.py also eliminates an undeclared requests dependency.