2020-09-22, 15:25–15:55, Room 2
Hand-crafted shaded relief is an effective yet tedious method to depict terrain on topographic maps. Convolutional neural networks trained with a large set of hand-drawn examples can create imagery that is strikingly similar to expressive hand-drawn shaded relief art.
Switzerland has a long tradition of crafting high-quality shaded relief representations for showing terrain on maps, but only a small number of experts still exist that are able to produce shaded relief art of high quality. There have been efforts to automate the procedure by applying variations of Lambert’s diffuse reflection law to digital elevation models, but up to now no computer-based approach exists that yields results matching the expressiveness of hand-crafted masterpieces. We explored different approaches and architectures of neural networks to replicate the style of hand-crafted shaded relief. The most promising approach is a modified deep U-Net architecture, trained on shaded relief imagery of the Swiss topographic map series of scales 1:25,000 and 1:200,000, respectively. Shaded relief images generated with this architecture not only exhibit the ability of neural networks to learn the hand-crafted style, but also remove unnecessary terrain details and locally adjust the light direction and emphasize large landforms, which are essential techniques used by cartographers for manual relief shading. A trained network generates “neural shadings” within seconds for large areas, enabling cartographers world-wide to augment their maps with high quality shaded relief computed from a digital elevation model. Hence, for this application deep learning is able to curate knowledge of the past that otherwise might eventually get lost.