Spatially resolved rainfall streamflow modeling in central Europe

Climate change increases the risk of disastrous floods and makes intelligent freshwater management an ever more important issue for society. A central prerequisite is the ability to accurately predict the water level in rivers from a range of predictors, mainly meteorological forecasts. The field of rainfall runoff modeling has seen neural network models surge in popularity over the past few years, but a lot of this early research on model design has been conducted on catchments with smaller size and a low degree of human impact to ensure optimal conditions. Here, we present a pipeline that extends the previous neural network approaches in order to better suit the requirements of larger catchments or those characterized by human activity. Unlike previous studies, we do not aggregate the inputs per catchment but train a neural network to predict local runoff spatially resolved on a regular grid. In a second stage, another neural network routes these quantities into and along entire river networks. The whole pipeline is trained end-To-end, exclusively on empirical data. We show that this architecture is able to capture spatial variation and model large catchments accurately, while increasing data efficiency. Furthermore, it offers the possibility of interpreting and influencing internal states due to its simple design. Our contribution helps to make neural networks more operation-ready in this field and opens up new possibilities to more explicitly account for human activity in the water cycle.