Difference between non-existent and non-available observations

[luukvdmeer]

Currently the pixel value nan (nodata) is used for three different cases:

• Pixels that are not real observations. These pixels are only present because a DataArray needs to be rectangular. When your spatial extent spans multiple orbits, some timestamps will have observations in areas where other timestamps don't, and vice versa. To make the whole array rectangular xarray creates pixels at the "missing observations" and fills them with nan. But of course these are not really observations.
• Pixels that should have an observation but its value is not available. These are invalid observations. If I understand it correctly it is defined in the data cubes metadata which values should be considered "invalid".
• Pixels that are removed by applying the filter verb.

The semantique package treats all nan values as if they don't exist. Hence, when applying the percentage reducer, it considers the "whole" to be the count of all non-nan pixels. When applying the first reducer, it returns the first non-nan pixel value. When evaluating the or operator, it returns nan when X = nan and Y = True (i.e. it preserves nan values in the input). The assign verb assigns new values only to non-nan pixels, the smooth verb only smoothes non-nan pixels.

We now use the same value for three cases with different semantics. We might need to consider using different values (but which??).