Using lon and lat as inputs to X when classifying heatwaves

[georgemilosh]

Actually it might be the case that better skill could be achieved if we took $X$ we have, i.e. $X = (t,z,s)$, and concatenate it with lat and lon (stacking): $X = (t,z,s,lat,lon)$. The CNN is normally has translation invariance which could be a problem when detecting features in a specific area of the map. By using explicitly fixed in time lat and lon we may break this invariance and extract useful information about the location of various features such as cyclones in the Atlantic for example.