Missing S2 tile in L4A classification about sen2agri-system HOT 10 CLOSED

It's not very clear from the screenshots, but we noticed some S2 tiles where the density of the acquisitions is much worse than on the neighboring ones. Can you compare the flags for 35NRH and its neighbors?

from sen2agri-system.

It looks like the density is lower - but given that there is both S2 and in-situ data - shouldn't there also be classified pixels?

Here's two screenshots comparing the flags:

and without 35NRH:

from sen2agri-system.

Can you check the values? There might be some scaling involved to fit the histogram of the image.

from sen2agri-system.

I'm not familiar with the actual info contained in the flag images, so I'll just post the histograms:

35NRH:

35NQH:

36NTN:

from sen2agri-system.

The four bands should be:

• the number of valid pixels across the time series
• the number of water pixels across the time series
• the number of snow pixels across the time series
• the number of other invalid pixels (cloud, no data, saturated)

So my theory doesn't seem to hold. I recall seeing this issue before, but it was in areas with worse coverage and / or a lot of clouds. I'm not sure what else might explain it.

from sen2agri-system.

Thanks @lnicola ... would be interesting what caused it, I didn't find any error msgs in the logs.

Anyways, I've re-run the classification for this tile and now everything seems to be complete. So I'm closing the issue.

from sen2agri-system.

I don't think it was an error, but the actual classifier output.

Anyways, I've re-run the classification for this tile and now everything seems to be complete.

Did you re-run it for the whole site, or just on that specific tile? If it was for the whole site with the same parameters and input products, that's interesting.

from sen2agri-system.

I don't think it was an error, but the actual classifier output.

But given that there's both reference and satellite data, how can an empty tiff file be a (correct) classifier output?

For the re-run, I've used the --tile-filter argument to restrict the classification, post-process and qualityflags steps to this specific tile. The strata, models, etc. was all from the previous run which was for the entire site.
The final quality metrics step was again done for the full site, just this time with the new tiff files for 35NRH replacing the previous ones.

from sen2agri-system.

But given that there's both reference and satellite data, how can an empty tiff file be a (correct) classifier output?

The classifier is somewhat resilient to noise, so the classification result doesn't have to match the training data exactly (there might be outliers, for example).

For the re-run, I've used the --tile-filter argument to restrict the classification, post-process and qualityflags steps to this specific tile. The strata, models, etc. was all from the previous run which was for the entire site.

The classification output should be (and usually is, as far as I know) deterministic, so.. that shouldn't have happened. I've no idea why, sorry.

from sen2agri-system.

That's why I think there must have been some kind of error. I'm quite familiar with Random Forest, but given there's a few reference data points, I wouldn't expect an output that has not a single crop pixel.

For now it's not a big issue, as far as this run is concerned, this seems to have been an isolated issue. But if it happens again, I would like to look a bit closer as to why that happens.

Thanks for now @lnicola

from sen2agri-system.