From my work in the Art of Feature Engineering book, I have learn to appreciate the value of standalone features that can incorporate domain knowledge into a problem.
As part of the #CleanMaltaAI hackathon Presagis team, I put together a small feature server that transforms Open Street Maps into a floating point feature for each point in a satellite image. This feature, therefore, injects structured information (maps) into a statistical/deep learning system.
The whole thing is packed so it can be run remotely through the included Flask server.
See the details in the GitHub repo for details.