Filter_bands() VS band()

openEO Platform
1

I’m trying (with python) to mask cloudy pixels from terrascope LAI product following the openEO cookbook.

No problem if I strictly follow the cookbook example use the .band() to select the classification layer and build a mask:

import openeo
# load data from openo
con  = openeo.connect("https://openeo.vito.be").authenticate_oidc(provider_id="egi")

# Load data cube from TERRASCOPE_S2_NDVI_V2 collection.
datacube = con.load_collection("TERRASCOPE_S2_LAI_V2",
                               spatial_extent={"west": 5.60, "south": 50.42, "east": 6.3, "north": 50.7},
                               temporal_extent=["2022-07-01", "2022-08-15"],
                               bands=["LAI_10M","SCENECLASSIFICATION_20M"])
# get classification band
SCL = datacube.band("SCENECLASSIFICATION_20M")
LAI = datacube.band("LAI_10M")
# masking
mask = ~ ((SCL == 4) | (SCL == 5))
LAI_masked = LAI.mask(mask)

But the .band() is not documented in the openEO hub and I tried to use .filterbands()… And I’m getting an error when using the mask I built with .filter_bands(), meaning that the two processes results in different outputs. Do you have an explanation?

# get classification band
SCL = datacube.filter_bands(bands = ["SCENECLASSIFICATION_20M"])
LAI = datacube.filter_bands(bands = ["LAI_10M"])

# masking
mask = ~ ((SCL == 4) | (SCL == 5))
#Error: OperatorException: Unsupported unary operator 'not' (band math mode=False)
2

band is a convenience function of the Python client and not a process implemented by the back-end. As such it is not listed on the openEO Hub. It also does a bit more than just filter_bands as it then allows the “easy” way to formulate band math in the client. The filter_bands just returns a new filtered data cube and as such, you can’t directly use it for math. What you could likely do is SCL = datacube.filter_bands(bands = "SCENECLASSIFICATION_20M"]).band("SCENECLASSIFICATION_20M")

3

Thanks!

I wandering how to guess that because both function are supposed to produce openeo.rest.datacube.DataCube objects?

4

In addition to what Matthias said, you could easily check how the first process graph looks like with:

import openeo
# load data from openo
con  = openeo.connect("https://openeo.vito.be").authenticate_oidc(provider_id="egi")

# Load data cube from TERRASCOPE_S2_NDVI_V2 collection.
datacube = con.load_collection("TERRASCOPE_S2_LAI_V2",
                               spatial_extent={"west": 5.60, "south": 50.42, "east": 6.3, "north": 50.7},
                               temporal_extent=["2022-07-01", "2022-08-15"],
                               bands=["LAI_10M","SCENECLASSIFICATION_20M"])
# get classification band
SCL = datacube.band("SCENECLASSIFICATION_20M")
LAI = datacube.band("LAI_10M")
# masking
mask = ~ ((SCL == 4) | (SCL == 5))
LAI_masked = LAI.mask(mask)
LAI_masked = LAI_masked.save_result(format='netCDF')
LAI_masked.create_job(title='LAI_masked_band')

which results in:

image
image791×693 20.6 KB

looking at the details, we can notice that the python client translates the lines:

SCL = datacube.band("SCENECLASSIFICATION_20M")
mask = ~ ((SCL == 4) | (SCL == 5))

into a reduce_dimension process, with the provided formula:

image
image1644×226 9.35 KB

Finally, the difference with using filter_bands is in the dimensions of the output datacube:
.bands() translates into a reduce_dimension + array_element, so the output datacube won’t have the bands dimension any more (has been reduced). Instead, filter_bands takes just one or some bands and keeps the bands dimension.
I guess that the Python client complains since it doesn’t know which of the bands has to use for the
mask = ~ ((SCL == 4) | (SCL == 5)) , since filter_bands output can have multiple bands.

@stefaan.lippens what do you think?

1 Like
5

Great explanations! Thanks :smile:

6

Indeed, as noted above, the .bands() method is convenience functionality (currently a feature only available in the openeo Python client, called “band math”) to build a reduce_dimension operation along the “band” dimension, using standard math operators. E.g.

red = cube.band("red")
green = cube.band("green")
result = 2 * red - 5 * green

will translate to something like (pseudo code):

reduce_dimension(
    dimension="bands",
    reducer=subtract(multiply(array_element(cube, "red"), 2), multiply(array_element(cube, "green"))))
)

Obviously, the math notation version is easier to understand.

.filter_bands() however is a standard openEO process to filter out specific bands (one or more) from a cube. It does not allow the same “band math” functionality, because it does not imply a reduce_dimension along the band dimension.
However, on the output of .filter_bands(), just like a normal DataCube, you still can apply mathematical operators. E.g.

cube = cube.filter_bands(["red", "green"])
cube = 2.5 * cube

This does not translate to reduce_dimension, but a “pixel-wise” apply, (scaling all pixel values with factor 2.5 in this case).

I hope this makes some things clear