Applying non-parametric Bayesian networks to estimate maximum daily river discharge: potential and challenges

Abstract. Non-parametric Bayesian networks (NPBNs) are graphical tools for statistical inference widely used reliability analysis and risk assessment present several advantages, such as the embedded uncertainty quantification limited computational time process. However, their implementation in hydrological studies is still scarce. Hence, to increase our understanding of applicability extend use hydrology, we explore potential NPBNs reproduce catchment-scale dynamics. Long-term data from 240 river catchments with contrasting climates across United States Catchment Attributes Meteorology Large-sample Studies (CAMELS) set will be actual means test utility descriptive models evaluate them predictive maximum daily discharge any given month. We analyse performance three networks, one unsaturated (hereafter UN-1), saturated SN-1), both defined only by hydro-meteorological variables bivariate correlations, network SN-C), consisting SN-1 including physical catchments' attributes. The results indicate that UN-1 suitable a positive dependence between precipitation discharge, while can also negative dependence. latter characteristics (tested via Kolmogorov–Smirnov statistic) have Nash–Sutcliffe efficiency (NSE) ≥0.5 ∼40 % analysed, receiving mainly winter located energy-limited regions at low moderate elevation. Further, SN-C network, based on similarity catchments, statistics ∼10 analysed. show once NPBN defined, it straightforward infer itself additional variables, i.e. going network. suggest considerable challenges defining NPBN, particularly predictions ungauged basins. These due discrepancies timescale different processes generating presence “memory” system, Gaussian-copula assumption modelling multivariate