The open_geotiff accessor method...

[brendancol]

For folks that don't know, xarray-spatial provides a GDAL-free open_geotiff function:

from xrspatial import open_geotiff

# scenario 1: directly use open_geotiff to create an xarray.DataArray
my_study_area_elevation = open_geotiff("elevation.tif")

xarray-spatial also provides an xr.DataArray / xr.Dataset accessor method .xrs which exposes the open_geotiff function ON an xarray.DataArray or xarray.Dataset object that can then serve as a template for loading additional data:

Simple example:

from xrspatial import open_geotiff

my_study_area_elevation = open_geotiff("elevation.tif")

# scenario 2: use open_geotiff through **accessor** method
my_landcover_data = my_study_area_elevation.xrs.open_geotiff("landcover.tif", unpack=True)  # equivalant to `mask_and_scale=True`
my_landcover_data = my_study_area_elevation.xrs.open_geotiff("landcover.tif", auto_reproject=True) # infers resampling param based on data but can be overridden
my_landcover_data = my_study_area_elevation.xrs.open_geotiff("landcover.tif", coregister=True) # tranform to match CRS, bounds, and shape of caller

A few things to note here:

• unpack parameter is analogous to mask_and_scale, and named because it aligns with the inverse operation, which is called pack which is easier to say than unmask_and_unscale.
• Then there is also, because we have metadata on the caller object, the data array or the data set, there is a parameter called auto_reproject, which will auto reproject the data into the, to match the caller's CRS.
• coregister indicates to return the output with matching shape, bounds, chunks and CRS of the caller object.

The vision here is that scaling, masking, reprojecting, and resampling are common enough tasks that it will be valuable to users to focus on making those steps easy from the beginning to speed up time to "analysis ready" arrays.

There are some difficult edge cases:

• I added in a warning in the scenario for:

huge_dask_array.xrs.open_geotiff("tiny.tif", coregister=True)

• datum transformation are not handled...

Would love to hear what people think.

[Melissari1997]

To clarify, this means we can read multiple TIFF files at once and stack them together into an xr.Dataset? And also align them so they have the same shape and CRS?
If so, this is a feature I would have used a lot when we made the switch from TIFF to Zarr—we had to read and stack many TIFF files together to generate an xr.Dataset to save in Zarr format

[brendancol]

@Melissari1997 Yes! The idea is essentially a convention that allows users to opt into using an existing xr.DataArray as the template for opening additional data.

something like:

import xarray as xr
from xrspatial import open_geotiff

# The first file becomes the template grid that everything else is snapped to...
study_area = open_geotiff("elevation.tif")

Then you could load the additional layers like:

layers = {
    "elevation": study_area,
    "landcover": study_area.xrs.open_geotiff("landcover.tif", coregister=True, resampling="nearest"),
    "ndvi": study_area.xrs.open_geotiff("ndvi.tif", coregister=True),
    "rainfall": study_area.xrs.open_geotiff("rainfall.tif", coregister=True)
}

study_area_ds = xr.Dataset(layers)

My hope is that coregister=True feels ergonomic. resampling can also be provided to override the resample method when coregistering.

[Melissari1997]

This is something I would use extensively for all use cases that rely on TIFF files!

[martindurant]

Can the geotiff loader be accessed through xr.open_[mf]dataset(engine=) instead of importing xrspatial explicitly?

[brendancol]

btw...akimbo looks sweeeeet

[martindurant]

It's a great idea searching for use cases ... Maybe the geo, IP or any other extensions I can think of will bring these. OTOH, you can get far with something like nested-pandas when you know just how deep or "awkward" your data will be.

[brendancol]

@martindurant ok so i got the engine backend configured and added the necessiary backend params to do the coregister...the spelling for the mfdataset currently looks like:

(xarray-spatial-everything) brendan@DESKTOP-7S2LU4R:~/geoenrichment/data/raw/LF2024_EVC_CONUS/Tif$ python
Python 3.14.2 | packaged by Anaconda, Inc. | (main, Dec 15 2025, 17:36:41) [GCC 14.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import numpy as np
>>> import xarray as xr
>>> arr = np.zeros((5,5), dtype=np.float64)
>>> template = xr.DataArray(arr, dims=("y", "x"), coords={"y": np.linspace(45.3, 44.7, 5), "x": np.linspace(-120.3, -11\
9.7, 5)}, attrs={"crs": 4326}, name="special-array")
>>> tif_glob = "L*.tif"   # I was testing on large landfire raster
>>> ds = xr.open_mfdataset(
...     tif_glob,
...     engine="xrspatial",
...     backend_kwargs={
...         "coregister": template,
...         "default_name": "landfire",
...     },
... )
>>> ds.landfire
<xarray.DataArray 'landfire' (y: 5, x: 5)> Size: 200B
dask.array<open_dataset-landfire, shape=(5, 5), dtype=float64, chunksize=(5, 5), chunktype=numpy.ndarray>
Coordinates:
  * y        (y) float64 40B 45.3 45.15 45.0 44.85 44.7
  * x        (x) float64 40B -120.3 -120.2 -120.0 -119.8 -119.7
Attributes:
    _xrspatial_geotiff_contract:  5
    georef_status:                full
    crs:                          GEOGCRS["WGS 84",ENSEMBLE["World Geodetic S...
    transform:                    (0.15, 0.0, -120.375, 0.0, -0.15, 45.375)
    gdal_metadata:                {('RESAMPLING', 0): 'NEAREST'}
    gdal_metadata_xml:            <GDALMetadata><Item name="RESAMPLING" sampl...
    extra_tags:                   [(305, 2, 192, 'GDAL 3.8.1e (3080114) (4.6....
    mask_and_scale_dtype:         int16
    nodata:                       nan
    masked_nodata:                True
    nodata_pixels_present:        False
    res:                          (0.15, 0.15)

[martindurant]

LGTM, I think this makes it easier to use for most, even if providing like= is unusual/

[brendancol]

Yes, like is unusual so I decided to instead overload coregister to take a template and removed the like parameter...thanks again for your input here.