Enhancing Water Level Estimates with DEM-derived Stream Geomorphometry

Accurate water level predictions are increasingly crucial for mitigating flood risks. Hydrological and hydrodynamic models provide water level predictions, but their accuracy depends on detailed information about stream cross-sections and floodplain topography, which are data that are difficult to obtain at larger scale, especially in regions with perennial river systems. Stream discharge is a variable that is more straightforward to predict by conventional hydrological models. However, the relationship between discharge and water level is complex, depending on cross-section geometry and channel roughness. Here machine learning models offer an alternative opportunity to predict water level by ingesting readily available topographic data derived from high-resolution digital elevation models in combination with simulated stream discharge, thereby skipping the need to explicitly define rating curves or to run complex hydrodynamic simulations. The idea is that stream discharge provides information about the temporal variability, whereas the topographic data provides static information in the cross-section geometry.  First, we present a method for extracting stream geomorphometry from a high-resolution (40 cm) digital elevation model in Denmark. The methodology is based on analyzing elevation changes along cross-sections throughout the entire Danish river network. Stream widths are estimated by identifying the most probable bank positions through a probabilistic count of all possible configurations within 100-meter stream reaches. The resulting dataset has been validated against 2,000 measured cross-sections along Danish rivers, showing similar spatial patterns across reach to river scales. Moreover, the slope and elevation of the water level as well as channel area and depth are derived from the high-resolution DEM for 100-meter stream reaches.Second, we present the development of a machine learning-based model that utilizes the derived stream geomorphometry in combination with stream discharge simulated by the National Hydrological Model of Denmark to predict daily stream water levels. Timeseries of daily stream water level of 40 gauging stations are used to train a Long Short Term Memory network. The results demonstrate that incorporating topography-derived information of mean water level and slope, stream channel width, area, and depth, enhance the accuracy of the water level estimates. Overall, our approach provides a versatile approach providing crucial information on flood risks that can easily be scaled up to national scale.