Evaluating Current Statistical and Dynamical Forecast Techniques for Seasonal United States Coastal Sea Level Prediction

The need for skillful seasonal prediction of coastal sea level anomalies has become increasingly evident as climate change has increased the risk of coastal flooding events. Aiming to improve our ability to forecast coastal inundation risk on seasonal and longer time scales, NOAA and NASA initiated the RISE project, a collaborative effort focused on developing and assessing novel dynamical and statistical forecast methods for coastal sea level and inundation risk for US coasts. This presentation is an outgrowth of that project, initially based on a pilot study of monthly sea level anomaly forecast skill assessed at two tide gauge stations, San Diego CA, and Charleston SC. In this study, we evaluate several current forecast systems -- NCAR Community Climate System Model Version 4 (CCSM4), GFDL Seamless System for Prediction and Earth System Research (SPEAR), and ECMWF Seasonal Forecast System 5 (SEAS5) -- by calculating deterministic and probabilistic skill from a few decades (1993-2015) of their retrospective forecasts (“hindcasts”) and for lead times of up to 6-9 months. Additionally, we examine potential local enhancement of hindcast skill by two post-processing downscaling techniques, an observationally-based multivariate linear regression and a hybrid dynamical model approach, using the adjoint model of the Estimating Circulation and Climate of the Ocean (ECCO) system forced by observed and model-predicted surface forcings.

We find that all these approaches face challenges stemming from whether the modeled sea surface height sufficiently represents observed local variations of coastal sea level, because of ocean model limitations and because of inadequacies in both model initialization and ensemble spread. Some of these issues also complicate the ability of the downscaling techniques to improve probabilistic skill, even though they do somewhat improve deterministic skill. In general, while deterministic hindcast skill is considerably higher for San Diego than Charleston, ensemble spread metrics such as forecast reliability and sharpness are mediocre for both locations. Additionally, evaluating how well any technique predicts seasonal coastal sea level variations is considerably complicated by the forced trend component and particularly how it is estimated, especially for Charleston; .essentially, skill assessment of US coastal sea level seasonal prediction is also a trend detection problem. Moreover, these results are largely matched by hindcasts from a Linear Inverse Model (LIM), a simple stochastically-forced linear prediction model constructed from observations, suggesting that substantial improvement still remains for coastal sea level prediction.