Digital Surface Model generation from satellite imagery is a difficult taskthat has been largely overlooked by the deep learning community. Stereoreconstruction techniques developed for terrestrial systems including selfdriving cars do not translate well to satellite imagery where image pairs varyconsiderably. In this work we present neural network tailored for DigitalSurface Model generation, a ground truthing and training scheme which maximizesavailable hardware, and we present a comparison to existing methods. Theresulting models are smooth, preserve boundaries, and enable furtherprocessing. This represents one of the first attempts at leveraging deeplearning in this domain.