Graph-Based Deep Learning on the Swiss River Network

Major European rivers have their sources in the Swiss Alps. Data from these rivers and their tributaries have been collected for decades with consistent quality. We use GIS data to extract the structure of each river and link this structure to 81 river water stations (that measure both water temperature and discharge). Since the water temperature of a river is strongly dependent on the air temperature, we also include 44 weather stations (which measure, for instance, air or soil temperature). Based on this large data corpus, we present in this paper a novel graph representing the water network of Switzerland. Our goal is to accelerate the research of the complex relationships at the (Swiss) water bodies. In particular, we present different graph-based pattern recognition tasks that can be solved on the novel water body graph. In a first evaluation, we use graph-based methods to solve two of these tasks, outperforming current state-of-the-art systems by several percentage points.