Scale transform changes aggregate results

[MSHelm]

Hey everyone,
as mentioned on Zulip I am using macsima data, where the macsima reader by default adds the Scale transformation to the coordinate system. I noticed that when I remove this my aggregation results are slightly different. This is most prominent for small signals, for larger signals it is not so pronounced. My original use case was for the mean intensity. For sum its very obvious that the absolute values change due to the transform, but the ratios stay similar.

Of course I understand that it makes sense to apply the transform before doing the aggregation, since they are intended for example for aligning my labels with the image. Still it was a bit unintuitive for me that this happens as a new user. Maybe this could be highlighted also a bit more in the tutorials. As requested below is a small example.

import spatialdata as sd
from spatialdata.datasets import blobs
from spatialdata.transformations import set_transformation, Scale

sdata = blobs()
# set scale transformation to new coordinate system. global coord system has Identity transform
scale_transform = Scale([0.1, 0.1], ("x", "y"))
set_transformation(sdata["blobs_image"], transformation=scale_transform, to_coordinate_system="scaled")
set_transformation(sdata["blobs_labels"], transformation=scale_transform, to_coordinate_system="scaled")

global_sum = sdata.aggregate(values="blobs_image", by="blobs_labels", target_coordinate_system="global")

print(global_sum["table"].X[:1])
#<Compressed Sparse Row sparse matrix of dtype 'float64'
#	with 3 stored elements and shape (1, 3)>
#  Coords	Values
#  (0, 0)	1309.3692551660652
#  (0, 1)	1587.8641823936478
#  (0, 2)	3125.1190857645483

scaled_sum = sdata.aggregate(values="blobs_image", by="blobs_labels", target_coordinate_system="scaled")

print(scaled_sum["table"].X[:1])
#<Compressed Sparse Row sparse matrix of dtype 'float64'
#	with 3 stored elements and shape (1, 3)>
#  Coords	Values
#  (0, 0)	12.758250581405642
#  (0, 1)	15.427204091295
#  (0, 2)	31.23271691509622


global_mean = sdata.aggregate(values="blobs_image", by="blobs_labels", agg_func="mean", target_coordinate_system="global")

print(global_mean["table"].X[:1])
#<Compressed Sparse Row sparse matrix of dtype 'float64'
#	with 3 stored elements and shape (1, 3)>
#  Coords	Values
#  (0, 0)	0.08696083251418378
#  (0, 1)	0.10545687603066001
#  (0, 2)	0.20755257260839133

scaled_mean = sdata.aggregate(values="blobs_image", by="blobs_labels", agg_func="mean", target_coordinate_system="scaled")

print(scaled_mean["table"].X[:1])
#<Compressed Sparse Row sparse matrix of dtype 'float64'
#	with 3 stored elements and shape (1, 3)>
#  Coords	Values
#  (0, 0)	0.08449172570467313
#  (0, 1)	0.10216691451188742
#  (0, 2)	0.2068391848681869

[LucaMarconato]

Thanks for reporting. I triaged this and this is due to the fact that scaling an image currently changes the resolution.

@melonora and I and working on improving the transformations to avoid this behavior, and rather scaling the xarray coordinates. This should take 1-2 months to be implemented.

On top of this there is a problem in the comparison here:

spatialdata/src/spatialdata/_core/operations/aggregate.py

Line 155 in aec5181

if not (by_transform == values_transform and isinstance(values_transform, Identity)):

• The two scales transformations are identical but the equality gives false. This should be fixed.

Current workaround:

• you can call rasterize before calling aggregate (and add the rasterized objects in the sdata object, or a new one, so to avoid having a scale transformations inside. Not ideal, but would work for now.