Issue with cropped `xarray` data causing `xarray`'s `.where()` to behave differently

[robbibt]

While testing #114, I've been running into a really confusing issue that seems to cause xarray's .where to change the output size of my arrays.

Usually, a.where(b) simply masks out values from a that are False in b, replacing them with NaN. This normally doesn't have any effect on the shape of a (assuming the default drop=False) - you get back an array of the orginal shape after applying the mask.

For example, I might want to rasterize a polygon, then apply this to my original data as a mask. This works fine - my masked data comes out the same size as my original data:

import odc.geo.xr
from odc.geo.geobox import GeoBox
from odc.geo import geom

# Create xarray
geobox = GeoBox.from_bbox((-10, -2, 5, 4), "EPSG:4326", resolution=0.2)
xx = odc.geo.xr.xr_zeros(geobox)
print(f"Original shape is: {xx.odc.geobox.shape}")

# Create polygon overlapping xarray
poly = geom.Geometry(
    {"type": "Polygon", "coordinates": (((-5, 3), (0, 2), (-2.5, -1)),)}, "EPSG:4326"
)

# Rasterize polygon
rasterized = odc.geo.xr.rasterize(
    poly=poly.to_crs(xx.odc.geobox.crs), how=xx.odc.geobox
)
print(f"Rasterized shape is: {rasterized.odc.geobox.shape}")

# Apply mask
xx_masked = xx.where(rasterized)
print(f"Masked shape is: {xx_masked.odc.geobox.shape}")

# Verify that original data and cropped data have same geobox
assert xx.odc.geobox == xx_masked.odc.geobox

Problem

However, if I modify the shape of my array using .isel() or similar, .where() starts to behave differently. This example is the same, except I select a subset of my array using .isel() before rasterizing my polygon into its new geobox.

The rasterization works fine - my rasterized polygon and clipped array have the same geobox and shape. However, when I attempt to combine them with xx_cropped.where(rasterized), I get out an array with a different geobox, including different resolution and shape:

import odc.geo.xr
from odc.geo.geobox import GeoBox
from odc.geo import geom

# Create xarray
geobox = GeoBox.from_bbox((-10, -2, 5, 4), "EPSG:4326", resolution=0.2)
xx = odc.geo.xr.xr_zeros(geobox)
print(f"Original shape is: {xx.odc.geobox.shape}")

# Create polygon overlapping xarray
poly = geom.Geometry(
    {"type": "Polygon", "coordinates": (((-5, 3), (0, 2), (-2.5, -1)),)}, "EPSG:4326"
)

# Select subset
xx_cropped = xx.isel(latitude=slice(10, 20), longitude=slice(20, 50))
print(f"Cropped shape is: {xx_cropped.odc.geobox.shape}")

# Rasterize polygon
rasterized = odc.geo.xr.rasterize(
    poly=poly.to_crs(xx_cropped.odc.geobox.crs), how=xx_cropped.odc.geobox
)
print(f"Rasterized shape is: {rasterized.odc.geobox.shape}")

# Verify that cropped and rasterized data have the same geobox
assert rasterized.odc.geobox == xx_cropped.odc.geobox

# Apply mask
xx_masked = xx_cropped.where(rasterized)
print(f"Masked shape is: {xx_masked.odc.geobox.shape}")

# Verify that masked data and cropped data have same geobox
assert xx_cropped.odc.geobox == xx_masked.odc.geobox

Am I doing something silly here? Or is there some interaction going on between xarray and odc-geo that is causing .where to behave differently on data that has previously subsetted with .isel()?

The confusing thing is that xx_cropped and rasterized seem to have completely identical pixel grids, i.e. same geobox and even the same spatial dim coordinates...

[robbibt]

Comparison of the arrays... somehow xx_cropped and rasterized combine to make xx_masked...

[Kirill888]

that would be due to floating point precision most likely.

In theory we should have an invariant like this

xr_coords(gbox)['y'][3:7].data == xr_coord(gbox[3:7, :])['y'].data

but this might not be the case due to numeric precision.

[Kirill888]

changing

- xx_masked = xx_cropped.where(rasterized)
+ xx_masked = xx_cropped.where(rasterized.data)

gives expected result as it ignores tiny differences in lon/lat:

rasterized.latitude.data - xx_cropped.latitude.data
>>> array([ 0.00000000e+00,  2.22044605e-16,  2.22044605e-16,  0.00000000e+00,
        0.00000000e+00, -2.22044605e-16,  0.00000000e+00,  0.00000000e+00,
       -2.22044605e-16,  0.00000000e+00])

[Kirill888]

in general, when using xarray.where() use plain numpy array that matches data dimensions to be sure to bypass coordinate matching logic.

[robbibt]

Thanks Kirill!