It seems to me that the xview training data is nearly useless as the distribution of pixels is just too different between xview and malawi data.

I am curious to learn what the top guys are doing