Use of Bayesian Networks to Combine USGS Surface Water Observations with Hydraulic Geometry Models towards Riverine Discharge Estimation

Abstract:

We have developed a probabilistic framework for riverine discharge estimation that describes more than 725,000 average daily streamflow observations from the U.S. Geological Survey’s Surface Water Field Measurement Program at 1999 sites located throughout the continental United States and collected since 2000. These data, combined with geomorphic parameter information obtained from the National Hydrography Dataset and the National Elevation Dataset, were used to train Bayesian Networks consisting of 9-15 parameters. Hydraulic geometry relationships were used to constrain power law coefficients such that observed time series representing the measured daily discharge could be correlated with synthetic time series governing hydraulic width, depth and velocity at each site. This approach allows the training set to represent natural flow conditions and not the somewhat irregular timing of the field surveys. Sensitivity analysis of network is presented, along with over 500 worldwide validation cases to demonstrate network accuracy when confidently constrained. Making discharge predictions that combine field data with empirical relations though the use of a probabilistic framework is advantageous in that it explicitly accounts for uncertainty in the interpretation of results. This allows objective assessment of the predictive performance when applied to qualitative decisions relevant to environmental managers.