zonal_aggregate_raster and zonal_aggregate_polygon may give unexpected/unusable results

In GitLab by @Huite on Nov 18, 2023, 17:38

I got this report from Jacco.

If the resolution is small enough in these methods, it will result in duplicate entries in the output:

imod.prepare.zonal_aggregate_raster(peilvlakken_shp, 'Id', da_1, 1.0, 'mean')
   Id       area  aggregated
0   1  1066795.0         1.0
1   2   279899.0         1.0
2   3  1162011.0         1.0
3   4  1491911.0         1.0
4   5   258044.0         1.0
5   6   419897.0         1.0
6   7    89131.0         1.0
7  11   593048.0         1.0

imod.prepare.zonal_aggregate_raster(peilvlakken_shp, 'Id', da_1, 0.5, 'mean')
   Id        area  aggregated
0   2    94655.00         1.0
1  11   294333.25         1.0
0   1  1066068.50         1.0
1   2   183242.75         1.0
2   3  1162725.75         1.0
3   4  1489183.75         1.0
4   5   257467.75         1.0
5   6   420808.75         1.0
6   7    89104.25         1.0
7  11   214502.00         1.0
0  11    59612.75         1.0
0  11    23599.50         1.0

The reason is relatively simple, internally the data is chunked and processed out-of-core using dask.

    raster_chunks, _, _ = _create_chunks(raster, resolution, chunksize)
    collection = [
        dask.delayed(_zonal_aggregate_raster)(path, column, resolution, chunk, method)
        for chunk in raster_chunks
    ]
    result = dask.compute(collection)[0]
    return pd.concat(result)

Obviously, if a polygon is present in more than one chunk, the ID will be present more than once in the result, which is then just concatenated.

This isn't too big a deal for methods such as sum or mean, because all the information is available to do another reduction afterwards.
But for something like mode, it won't work, since that requires ALL values.

Start with zonal_aggregate_raster and zonal_aggregate_polygon, then read the shown _create_chunks, _zonal_aggregate_raster, and dask.compute flow. Reproduce the duplicate IDs with the supplied resolution examples and determine how chunk results should be reduced, including mode. Done means polygons spanning chunks produce one usable result with correct aggregation.