Neural networks for generalized relief shading, contour lines and coastlines

Artificial neural networks excel at analysing and transforming images and other raster fields. We apply U-nets, a type of convolutional neural network (Ronneberger et al., 2015) to cartographic generalization problems. We discuss three applications to derive generalized terrain representations from geospatial raster data (Guilbert et al., 2014). In the first application, we create generalized shaded relief; the neural network directly derives shaded relief images from digital elevation models. In the second application, we create generalized contour lines; the neural network generalizes the digital elevation model source to an intermediate elevation model, and generalized vector contour lines are derived from this intermediated elevation model with a contouring algorithm. In the third application, we create generalized coastlines with a neural network from a raster field that encodes the occurrence of surface water. The network produces a black-and-white coastline image, attuned to the continental landmasses and islands, which are then delineated as vector lines. We observe that high-quality reference data is required for training neural networks, and identify two complementary approaches to control the amount of generalization with neural networks that apply to the three described applications.