Performance of data-driven approaches for estimating flow hydrographs from rainfall hyetographs in small mountain catchments.

Data driven algorithms have been largely proven to be accurate tools for modelling many hydrological variables including aggregated river flows. Many studies have tested the suitability of a wide range of data-driven algorithms for predicting the recorded flows with times-steps ranging from a few minutes to monthly or even seasonal observations fed on a wide variety of inputs. They existing works often achieve brilliant performance indicators. In this work we pay our attention to a well-known hydrological process which is the flow hydrograph generation from rainfall hyetographs based on the mass conservation law within the catchment. Our assumption is that given many different physically based theories can provide accurate estimates of the expected flow hydrograph just providing the recorded hyetograph and a set of physical parameters of the catchment, data-driven approaches should also be able to successfully estimate the flow recorded hydrographs. For testing that hypothesis, we have selected two small mountain catchments (rivers Aragón in Canfranc and Valira Oriente in Andorra catchments in the Pirineos mountains in Spain and Andorra) easily parametrizable with no water depletion. We have checked the performance of different data-driven algorithms for predicting the 15-minutes recorded hydrographs fed on 15-minutes rainfall records and the set of physical variables involved in the Green-Ampt infiltration model. Over this process we have faced several issues and observed the data-driven algorithms are unable to provide the performance indicators commonly achieved in the published works.