I think it's using the contour lines on the map, not the villages and roads. The paper mentions the training data is contour maps and relief maps of the same area.