A data-driven approach to evaluate the importance of terrain features for hydrological modelling in flood warnings

Denmark’s national river flood forecasting system employs several different hydrological models for predicting river discharge estimates. These are used by duty meteorologists at the Danish Meteorological Institute (DMI) to issue flood warnings. In this study we explore two of these models: a Long Short-Term Memory network (DK-LSTM) and a conceptual hydrological model with the software HYPE (DK-HYPE). From operational experience, we suspect that both models have structural deficiencies related to lack of topographically driven processes. We therefore apply a dual-model approach to explore the potential in processing and feeding more detailed terrain description for forecasting high river flows in Denmark.

The LSTM model is trained primarily based on the CAMELS data set for Denmark. CAMELS data sets are widely used and are becoming a standard, recognized data set for training and running data-driven models. The current CAMELS data sets contain simple statistical description of terrain features, like catchment-averaged mean, min and max values of elevation above mean sea level and terrain slope. The HYPE model is based on the concept of hydrological response units (HRUs) but the default implementation in HYPE only delineates HRUs based on soil and land use information. Experience from the development of the two national models (DK-LSTM and DK-HYPE) indicates that catchments with distinct topological characteristics can exhibit markedly different hydrological responses that are not captured by simple catchment averages of DEM properties.

We perform detailed raster-based representations of terrain indices like HAND, rDune and TWI across Denmark. We then test multiple ways of processing the indices and summarizing the distribution of values within each sub-catchment into catchment attributes that the LSTM model can use as inputs. The relative importance of various terrain indices in DK-LSTM for high-flow predictions are then evaluated, and this information is used to redesign HRU delineation in the DK-HYPE. This enables the DK-HYPE setup to calibrate hydrological processes with terrain information. Our findings show which terrain indices, and therefore which topographic properties, that provide most benefit for predictive performance in high river discharge events relevant for flood warning applications.