cf.Field.collapse() fails when computing a volume mean where the volume cell measure is time-dependant

[dlrhodson]

Hi All,

If I try to compute a volume mean using cf.Field.collapse() where the cell measure is time dependant (sometimes true for ocean fields), cf.Field.collapse() throws an error.

mean=data.collapse('volume: mean',weights='volume')

Traceback (most recent call last):
  File "/gws/ssde/j25b/terrafirma/monitoring/references/cf_bug/./cf_collapse_test.py", line 7, in <module>
    mean=data.collapse('volume: mean',weights='volume')
  File "/home/users/dlrhodso/miniforge3/envs/cf20/lib/python3.14/site-packages/cf/decorators.py", line 71, in precede_with_kwarg_deprecation_check
    operation_method_result = operation_method(self, *args, **kwargs)
  File "/home/users/dlrhodso/miniforge3/envs/cf20/lib/python3.14/site-packages/cfdm/decorators.py", line 171, in verbose_override_wrapper
    return method_with_verbose_kwarg(*args, **kwargs)
  File "/home/users/dlrhodso/miniforge3/envs/cf20/lib/python3.14/site-packages/cf/field.py", line 7470, in collapse
    d_weights = f.weights(
        weights,
    ...<5 lines>...
        great_circle=great_circle,
    )
  File "/home/users/dlrhodso/miniforge3/envs/cf20/lib/python3.14/site-packages/cf/field.py", line 3488, in weights
    Weights.cell_measure(
    ~~~~~~~~~~~~~~~~~~~~^
        self,
        ^^^^^
    ...<4 lines>...
        auto=False,
        ^^^^^^^^^^^
    )
    ^
  File "/home/users/dlrhodso/miniforge3/envs/cf20/lib/python3.14/site-packages/cf/weights.py", line 1553, in cell_measure
    clm.squeeze(iaxes, inplace=True)
    ~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^
  File "/home/users/dlrhodso/miniforge3/envs/cf20/lib/python3.14/site-packages/cfdm/decorators.py", line 44, in inplace_wrapper
    processed_copy = operation_method(self, *args, **kwargs)
  File "/home/users/dlrhodso/miniforge3/envs/cf20/lib/python3.14/site-packages/cfdm/data/data.py", line 7178, in squeeze
    raise ValueError(
    ...<2 lines>...
    )
ValueError: Can't squeeze axis in position 1 from data with shape (1, 75, 332, 362): Axis size is greater than 1

cf.environment(paths=False)
Platform: Linux-5.14.0-611.54.1.el9_7.x86_64-x86_64-with-glibc2.34
Python: 3.14.4
packaging: 26.2
numpy: 2.4.3
cfdm.core: 1.13.1.0
udunits2 library: /home/users/dlrhodso/miniforge3/envs/cf20/lib/libudunits2.so.0
HDF5 library: 2.1.0
netcdf library: 4.10.0
netCDF4: 1.7.4
h5netcdf: 1.8.1
h5py: 3.16.0
pyfive: 1.1.2
zarr: 3.2.1
fsspec: 2026.4.0
scipy: 1.17.1
dask: 2026.3.0
distributed: 2026.3.0
cftime: 1.6.5
cfunits: 3.3.7
cfdm: 1.13.1.0
esmpy/ESMF: 8.9.1
psutil: 7.2.2
matplotlib: 3.10.9
activestorage: not available
cartopy: 0.25.0
cfplot: not available
cf: 3.20.0

This bug occurs due to a line in the Weights.cell_meaure() function. The line:
iaxes = [clm_axes0.index(axis) for axis in clm_axes]
is intended to select redundant axes to be squeezed, but instead selects the non-redundant axes. This part of the code is only ever reached if the cell_measure has at least one axis of size 1. This occurs if cell_measure is time dependant and we are trying to collapse a field with a single time point.
What this line should do is to select all the axes in clm_axes0 that are not in clm_axes. This line:
iaxes = [clm_axes0.index(axis) for axis in (set(clm_axes0)-set(clm_axes))]
does this. Replacing with this line fixes the bug.

Attached is a zip file with a demonstration of the bug.

cf_bug.zip

[beanscg]

I checked current main, and the likely narrow spot is in the cell-measure axis reduction path:

• cf/weights.py:1533-1537 builds clm_axes by keeping only cell-measure axes whose data shape is greater than 1.
• When clm_axes != clm_axes0, cf/weights.py:1551-1553 computes iaxes from the retained axes and passes them to clm.squeeze(...).
• For a time-dependent volume measure with a singleton time axis, that looks like it may try to squeeze the retained non-singleton axes rather than the singleton axis that was removed from clm_axes.
• I did not see an open PR that appears to cover this issue.

A focused regression could use a volume cell measure with shape like (1, z, y, x) and assert the singleton time axis is removed while the spatial axes remain in the weight key/data.