The value of spatially-differentiated  catchments descriptors for predicting characteristics of hydrological events in German catchments

Abstract

Catchments descriptors are widely used in hydrological science to infer dominant hydrological processes, identify, and transfer information across catchments and scales. However, persistent use of descriptors aggregated as spatially-lumped values (i.e., catchment averages), without considering their spatial variability within catchments might hamper the efficiency of these tasks. In this study, we use interpretable machine learning to investigate the value of topographically enhanced catchment descriptors (i.e., weighting them using distance to outlet, distance and height to the nearest drainage and stream order) belonging to seven distinct categories (i.e., climate, topography, land use, geology, hydrogeology, soil physical properties, and soil water properties) for predicting mean values, variability and seasonality characteristics of hydrological droughts and runoff events occurred in 401 German catchments in the period 1979-2002. We found that the spatially-differentiated catchment descriptors aggregated with topographical enhancing are able to predict droughts and runoff events characteristics more accurately than the lumped descriptors. The improvement is particularly promising for prediction of runoff event characteristics. Particularly, descriptors aggregated using height above the nearest drainage and stream order are essential for accurate prediction of variability of runoff events characteristics, while the proximity to the stream and to the outlet are more relevant for predicting their seasonality. In case of droughts, the descriptors weighted by the proximity to the stream improve the predictions of the variability and seasonality of duration and severity (i.e., deficit volume) of hydrological droughts. Moreover, we show that spatially-differentiated aggregation has the potential to identify the importance of descriptors that appeared irrelevant when aggregated in lumped way, particularly shading a light on the role of mean annual potential evapotranspiration and forest land cover descriptors for the prediction of mean values and seasonality of time scale of runoff events, and the role of groundwater yield and wetland land cover to predict the variability of time rise of runoff events. Our study highlights that development of the methods for spatially-differentiated aggregation has potential to disentangle the effects of different physio-geographical controls on event response in different catchments and to improve its predictability in ungauged locations.