Probabilistic fusion of spatio-temporal data to estimate stream flow via Bayesian networks

Stream flow directly impacts flooding and transport of sediments and pollutants (e.g. NO₃) in watershed systems, and hence knowledge of current and future flow is important. Unfortunately, spatially dense networks of in situ stream flow meters are not generally available and would be prohibitively expensive to deploy and maintain. It is known that surface morphology, land cover, and rainfall can impact stream flow through time, and these quantities can be measured using remote sensing techniques. Physically-based hydrologic models have been used to estimate flow using such data, but such approaches are limited because severe assumptions often have to be made for many input values. We present an alternative approach based upon a spatio-temporal Bayesian network (STBN) solution and compare its performance to methods based on physical models and more approximate probabilistic models. We show that the STBN method is able to capture the non-stationarity of hydrologic processes that contribute to streamflow in space and time while also limiting the computational requirements.