Observation-based bankfull discharge estimates to improve global flood models

The conveyance capacity of rivers is a key uncertainty in regional and global flood models. Most models resort to assumptions of uniform discharge recurrence of 1-2 years, using modelled discharge. While this assumption may hold on average, reach-scale bankfull discharge has been shown to vary significantly at the global scale. To improve this key boundary condition in large-scale hydrodynamic models, we have coupled emerging understanding of the hydrological and geomorphological drivers of bankfull discharge with recent advances in remote sensing products and machine learning. Using measured bankfull discharge values derived from stage-discharge and width-discharge relationships as reference, we construct a data-driven model to estimate bankfull discharge globally at the reach scale (30m centreline pixels and sub-kilometre vector reaches). Various remote sensing products are used as predictor variables that pertain to either catchment-wide or reach-specific attributes. This includes river geometry and floodplain metrics derived from Landsat water masks that have also been used to construct the underlying river network. This novel river network was designed to be as DEM-independent as possible, allowing for multi-thread channels, bifurcations and canals. Early results indicate good agreement between predicted and independent reference values.

The new dataset will be used to improve the parametrisation of a state-of-the-art global flood model as part of the EvoFlood research project (NERC, UK), but is also expected to be useful for other hydrological and hydrodynamic models as well as investigations at regional to global scales.