"to_scl_dilation_mask" does not work as intended with filter_spatial

[GriffinBabe]

When executing the "to_scl_dilation_mask" process and using the filter_spatial(geometry), undesired results arrive. The array contains NaN, 1.0 and very few 0.0 values.

Here is how the extracted scl dilated mask look like, other dates from the datacube yield the same result

The process graph is the following:

I managed to reproduce the issue with this minimal example:

import openeo

connection = openeo.connect('http://openeo-staging.dataspace.copernicus.eu')
connection.authenticate_oidc()

cube = connection.load_collection(
    'SENTINEL2_L2A',
    bands=['B04', 'B03', 'B02', 'SCL'],
    temporal_extent=["2020-01-01", "2020-06-01"],
)

geometry = connection.load_url(
    'https://artifactory.vgt.vito.be/artifactory/auxdata-public/gfmap-temp/openeogfmap_dataframe_0.parquet',
    format="Parquet"
)

cube = cube.filter_spatial(geometry)

scl_mask = cube.process(
        "to_scl_dilation_mask",
        data=cube.filter_bands(["SCL"]),
        scl_band_name="SCL",
        kernel1_size=17,  # 17px dilation on a 20m layer
        kernel2_size=77,   # 77px dilation on a 20m layer
        mask1_values=[2, 4, 5, 6, 7],
        mask2_values=[3, 8, 9, 10, 11],
        erosion_kernel_size=3
).rename_labels("bands", ["SCL_DILATED_MASK"])

cube = cube.merge_cubes(scl_mask)
# cube = cube.linear_scale_range(0, 65534, 0, 65534)


job = cube.create_job(
    out_format='NetCDF',
    title="To_scl_dilation",
    job_options = {
        "executor-memory": "5G",
        "executor-memoryOverhead": "2G",
    }
)

job.start_and_wait()

[GriffinBabe]

• Upload new geometries on static link with 1 bbox and multiple bbox
• Make the test before merging the cubes
• And then with the merging

[GriffinBabe]

@soxofaan
I uploaded the geoparquet files with one and three locations in the artifactory
https://artifactory.vgt.vito.be/artifactory/auxdata-public/gfmap-temp/openeogfmap_dataframe_singleloc.parquet
https://artifactory.vgt.vito.be/artifactory/auxdata-public/gfmap-temp/openeogfmap_dataframe_threelocs.parquet

One location, non-merged (cdse-staging j-240311d26137423897c1c73ab3301587)
So far I tested the scl dilated mask without merging it with the optical bands and it seems to work fine, although there are artifacts on the borders due to the convolution so not sure how can this be useful for WorldCereal project:

One location, merged (cdse-staging j-240311ad92974f4b828733e8a3717a13)
Here is the result after we merge the mask with the optical bands. Results are shifted:

Three locations, non-merged

Three locations, merged

[jdries]

FYI, the logs do show that a pixel buffer is applied:

 'pixel_buffer': [38.5, 38.5]}
Loading Sentinel2 with params DataCubeParameters(256, {}, FloatingLayoutScheme, ByDay, 6, None, NearestNeighbor, 38.5, 38.5)

The fact that there's still an edge effect means it's either not applied properly in this case, or not large enough.

[soxofaan]

FYI, as discussed in our meeting earlier this week:

• docs: to_scl_dilation_mask: document resolution of kernel/erosion? openeo-python-driver#271

[jdries]

Reassigned in Emile's absence, will look for unit test.

[jdries]

@GriffinBabe I've been trying to reproduce this. I'm focusing on the case where there are multiple patches, and using sample_by_feature.
In that case, there does not seem to be a border effect, both with and without merging optical bands.
I've used your geoparquet with 3 samples, focusing on date range: ("2020-03-29","2020-04-06")

Can you also try again, to help figure out what part leads to these artifacts?

[GriffinBabe]

@jdries
Indeed, when applying sample_by_feature it works (I didn't applied it previously as I was told by Stefaan to not take this into account for the moment)

There is still the issue with the border though. You said previously that I'm applying a pixel buffer of a certain size:

FYI, the logs do show that a pixel buffer is applied:

 'pixel_buffer': [38.5, 38.5]}
Loading Sentinel2 with params DataCubeParameters(256, {}, FloatingLayoutScheme, ByDay, 6, None, NearestNeighbor, 38.5, 38.5)

The fact that there's still an edge effect means it's either not applied properly in this case, or not large enough.

Is this related to the size of the polygons? I suppose this is, because that's the only moment in the pipeline where I am choosing a buffer, but the polygon size is 64x64 pixels, so it should be 32.0, not 38.5

What I am getting so far:

[jdries]

@GriffinBabe can you share the code again that produces this? Because I don't get the border.

The buffer is calculated based on the largest kernel size, so that the convolution is applied correctly. This is what normally should prevent border effects. The value of 38.5 is correct for a kernel size of 77.

[GriffinBabe]

This is the code:

import openeo

connection = openeo.connect('http://openeo-staging.dataspace.copernicus.eu')
connection.authenticate_oidc()

cube = connection.load_collection(
    'SENTINEL2_L2A',
    bands=['B04', 'B03', 'B02', 'SCL'],
    temporal_extent=["2020-01-01", "2020-06-01"],
)

geometry = connection.load_url(
    'https://artifactory.vgt.vito.be/artifactory/auxdata-public/gfmap-temp/openeogfmap_dataframe_threelocs.parquet',
    format="Parquet"
)

cube = cube.filter_spatial(geometry)

scl_mask = cube.process(
        "to_scl_dilation_mask",
        data=cube.filter_bands(["SCL"]),
        scl_band_name="SCL",
        kernel1_size=17,  # 17px dilation on a 20m layer
        kernel2_size=77,   # 77px dilation on a 20m layer
        mask1_values=[2, 4, 5, 6, 7],
        mask2_values=[3, 8, 9, 10, 11],
        erosion_kernel_size=3
).rename_labels("bands", ["SCL_DILATED_MASK"])

cube = cube.merge_cubes(scl_mask)
# cube = scl_mask
# cube = cube.linear_scale_range(0, 65534, 0, 65534)


job = cube.create_job(
    out_format='NetCDF',
    title="To_scl_dilation",
    job_options = {
        "executor-memory": "5G",
        "executor-memoryOverhead": "2G",
    },
    sample_by_feature=True,
)

job.start_and_wait()

[jdries]

Thanks, now I spotted it: you do the filter_spatial before to_mask_scl_dilation. Problem is that this process actually needs neighbouring pixels, but filter_spatial will remove them, so output is no longer correct.
Solution is to simply move filter_spatial after mask dilation process, that seems to work for me.