Improve documentation of SpaceNet DataModule

[nilsleh]

This should to be more clearly documented in the docs init. From the docs one expects building segmentation to be a binary classification task, but that throws errors then during training with the updated values. I spend an hour and was close to opening a Kornia issue, because I couldn't figure out why a normalization transform would change the values of the mask.

https://github.com/microsoft/torchgeo/blob/ba0680a98bb0229ac3cf79556eb1d956f326ae78/torchgeo/datamodules/spacenet.py#L86

[adamjstewart]

Want to submit a PR for this? Seems simple.

[calebrob6]

Oh wow I just went through this exact same thing last night -- was losing my mind for a bit. Was going to open an issue today. The +1 doesn't make sense as in the SN2 data, at least, as there are all black areas that map to class 0 already.