Machine Learning in Assessing the Performance of Hydrological Models

Machine learning has been employed successfully as a tool virtually in every scientific and technological field. In hydrology, machine learning models first appeared as simple feed-forward networks that were used for short-term forecasting, and have evolved into complex models that can take into account even the static features of catchments, imitating the hydrological experience. Recent studies have found machine learning models to be robust and efficient, frequently outperforming the standard hydrological models (both conceptual and physically based). However, and despite some recent efforts, the results of the machine learning models require significant effort to interpret and derive inferences. Furthermore, all successful applications of machine learning in hydrology are based on networks of fairly complex topology that require significant computational power and CPU time to train. For these reasons, the value of the standard hydrological models remains indisputable. In this study, we suggest employing machine learning models not as a substitute for hydrological models, but as an independent tool to assess their performance. We argue that this approach can help to unveil the anomalies in catchment data that do not fit in the employed hydrological model structure or configuration, and to deal with them without compromising the understanding of the underlying physical processes