That's actually a great idea! I wonder what kind of feature size would be needed though - TESSERA's embeddings are at a 10 metre resolution so for larger structures you might need some kind of spatial aggregation.