Using 'nearest' when having NaN

[rodrigo-j-goncalves]

I have a large sparse matrix of surface temperature values, indexed by latitude and longitude. The grid is irregular and has lots of NaN values (actually, most values are NaN).

I need to inspect the 2D matrix for temperature values at specific locations (i.e., arbitrary latitudes and longitudes).
Say I need the temperature value nearest to latitude = -24.5 and longitude = 56.3
Thus I thought I'd use something like

mylocationLat = -24.5
mylocationLon = 56.3
xarray.Dataset.sel(latitude = mylocationLat, longitude = mylocationLon, method = 'nearest')

If I do that, I almost always get a NaN (except that I am very lucky and use lat, lon really close to a non-NaN value in my matrix). And I guess this is not wrong, because my matrix actually has a cell 'near' my interest point... it's just that the cell has no temperature value on it. Thus it is not of much use to my case.

Right now I'm using a kind of convoluted approach, which uses slices and basically progressively increases the search intervals (i.e., min and max of the slice range) until I get a non-NaN value.

It would be interesting if I could tell xarray that I am looking for the nearest non-NaN value of my temperature variable. Makes sense? So, is there a way that I can tell xarray that the .sel(method = 'nearest') to keep looking until it finds a non-NaN data value? Alternative, could you suggest how (if possible) to mask all the NaN of my temperature values so sel(method = 'nearest') keeps looking until it finds a non-NaN value?