issues

What happened: I was computing an objective function between a remotely sensed snow covered area dataset and a simulated snow covered area dataset and every time I ran the code I got a different answer. I narrowed the issue down to when the monthly mean is taken of an image stack (geotiffs here) and then the mean is taken of the monthly values.

Maybe there is something about using geotiffs in this way that I am missing? The issue also occurs when I use the NaNs in the geotiffs to insert NaNs in the proper places in the simulation output. When I don't "cascade" the NaNs, the issue goes away in the simulated data but it always persists in the geotiffs.

What you expected to happen: I was expecting the mean of means output to be the same ever iteration as the same image stack was used. Minimal Complete Verifiable Example: I can't share the image stack at this time, but here is a shortened version of the code. ```python test=[] for i in tqdm.notebook.tqdm(range(2)): # load fSCA and assign dims ls = xr.open_mfdataset(files, parallel=True, chunks=dict(time=1,band=1, x=2500, y=2500), preprocess=add_time_coord) ls['time'] = dates ls = ls.rename_vars(dict(band_data='fsca'))

# threshold fSCA to SCA and cascade NaNs
ls['sca'] = xr.where(ls.fsca >= 150, 1, 0) # compute sca from fsca

test.append(ls.sca.groupby("time.month").mean().compute()) # monthly images

np.sqrt(np.mean(np.square((test[0].mean(('month'))-test[1].mean(('month'))).values)))

0.004654084555997707 ``` I expected the value here to be zero. Anything else we need to know?: The geotiffs have a no data value of -999 which is decoded to np.NaN

Environment: