This document discusses using graph convolutional networks to learn spatial relations from geographic data represented as graphs. It presents experiments on (1) detecting alignments between linear geographic features and (2) selecting roads from a generalized network that should be kept or erased at a target map scale. For alignment detection, proximity criteria were used to modify the graph and convolutional networks with attribute features predicted alignments with 72% accuracy. For road selection, features like length and importance were used and the method classified 72% of sections correctly with 73% precision and 83% recall. The document concludes graph convolutional networks show promise for learning spatial relations from graph-represented geographic data.