OP links to https://ai.facebook.com/blog/mapping-roads-through-deep-lear... which says that it's D-LinkNet specifically: http://openaccess.thecvf.com/content_cvpr_2018_workshops/pap... (more or less a Unet).
A huge amount of landcover segmentation in remote sensing still relies on simple models - either linear regression (thresholds) or classical machine learning like random forests or SVMs. For a lot of cases, these techniques will get you 90% of the way and it's very rare to have ground truth data that is accurate enough that you can measure the difference with any real degree of confidence.
A big problem in the field is the lack of good (public) ground truth. There's so little hand labelled data to work with that without humans in the loop it's extremely difficult to validate the results meaningfully (unless you have an army of staff to do it). With something like roads you could also have heuristics about what a road looks like and where it goes (e.g. it's a continuous thin line), which can help condition things.
I've seen a lot of papers which are applying deep learning for semantic segmentation for satellite mapping, but they evaluate on very limited datasets, they attempt to regress to simpler models without realising it (e.g. trying to predict a linear model), or they leak train and test data and report amazing results because they randomly split data from the same region.
I'm not saying that convnets aren't better than simpler models, but particularly for satellite imaging I'd take them with a pinch of salt and see what the improvement from a baseline method is. If you look at a random sampling of papers from the DeepGlobe competition, almost none of them provide the results from e.g. a cheap linear SVM.
Fun side note - several existing "famous" datasets generalise poorly to the developing world because most of the imagery is from the developed world (and even more specifically the West) and infrastructure looks totally different.
Have a look at mnist classification using a linear SVM, for example.