Performance of the US Navy's RELO Ensemble, NCOM, HYCOM During the Period of GLAD At-Sea Experiment in the Gulf of Mexico.

摘要

A suite of real-time ocean model forecasts was carried out successfully at NRL to provide modeling support and guidance to the CARTHE GLAD at-sea experiment during summer 2012. The forecast systems include two RELO ensembles and three single models using NCOM and HYCOM with different resolutions. All of these forecast outputs are archived and made available on web servers for the CARTHE scientists. The detailed descriptions of these forecast systems and the products presented in this paper provide a much-needed background to the scientists in CARTHE and others who will use our forecasts and GLAD drifter observations to do further research after the future public release of the CARTHE GLAD data. A calibrated ensemble system with enhanced spread and reliability is proposed in this project. It is found that this calibrated ensemble outperforms the un-calibrated ensemble in terms of quantitative forecasting accuracy, skill and reliability for all the variables and observation spaces we have evaluated. The metrics used include RMS error, anomaly correlation, spread-reliability and Talagrand rank histogram. Both ensembles are compared with three single-model forecasts with NCOM and HYCOM with different resolutions. The advantages of ensembles are demonstrated. RELO ensembles have been applied to Lagrangian trajectory prediction, and it is demonstrated that either ensemble can provide valuable uncertainty information in addition to predicting the particle trajectory with highest probability in comparison with a single ocean model forecast. The calibrated ensemble with more reliability is able to capture some trajectories in different, even opposite directions which are missed by the un-calibrated ensemble. When the ensembles are applied to computing the LCS (Lagrangian Coherent Structure), the uncertainties of the LCSs, which cannot be estimated from a single model forecast, are identified. Another finding is that the LCS depends on the model resolution.