A comparison of robust metamodels for the uncertainty quantification (UQ) of New York Harbor oceanographic data
M C Kerman, Wei Jiang, A. F. Blumberg, Samuel E. Buttrey · Journal of Operational Oceanography · 2008
This research demonstrates an innovative technique for the uncertainty quantification (UQ) of geophysical data using robust metamodels. Using salinity data supplied by the New York Harbor Observing and Prediction System (NYHOPS), five metrics are compared for seven different UQ methods. Metrics include covariance, correlation, mean error, mean absolute error, and mean absolute percent error. The seven UQ methods include a compromised data set (as the control case), two time-series models (moving average and auto-regression), two regression models (LOESS regression and quantile regression), Bayesian Gaussian Process modelling, and Kalman filtering. Upon examining the metrics computed for a set of 35 random sensors (out of the population of 735 sensors) over a five day period, the quantile regression metamodel is found to be superior to the other methods.