Verification of NCEP–GFS Ensemble Forecasts of Landfalling Atmospheric Rivers during Winter 2016–2017

Atmospheric Rivers (ARs) are characterized by long, narrow corridors of enhanced integrated water vapor (IWV) and integrated vapor transport (IVT) located in the warm sector of midlatitude extratropical cyclones. ARs are well-known contributors to high-impact weather events (e.g., precipitation extremes, floods, landslides) and water resources across Western U.S. The ability to provide advanced and accurate forecasts of ARs are key to a variety of impact-based decision support services and applications such as forecast-informed reservoir operations (FIRO; Ralph et al. 2014, JCWRE).

An AR Landfall Tool derived from a “probability-over-threshold” (P-O-T) IVT magnitude created by Cordeira et al. (2017, BAMS) illustrates timing, duration, and intensity of landfalling ARs along U.S. West Coast in NCEP Global Ensemble Forecast System (GEFS) forecasts. Anecdotal evidence indicates that this tool and the NCEP GEFS can accurately predict AR landfalls 6 to 8 days in advance, with some events predictable out to 10 or more days. The purpose of this presentation is to illustrate the AR-related skill of the NCEP GEFS and AR Landfall Tool during winter 2016–2017 along the U.S. West Coast.

Landfalling ARs during winter 2016–2017 contained an average lead-time predictability of ~6.25 days when forecasting ARs with an ensemble P-O-T IVT magnitude ≥250 kg m^–1 s^–1 exceeded and stayed above 50%. Landfalling ARs contained lead-time predictability >8–10 days in mid October, early January, and early February, while others contained lead-time predictability <5 days in March and April. More intense events appear to contain longer lead-time predictability. The GEFS AR Landfall Tool produced reliable forecasts at all leads between 1–16 days: A P-O-T of 50% verified ~66% of time out to 11 days, whereas a P-O-T of 70% verified ~70% of time out to 9 days. Receiver operator characteristic curves illustrate that the probability of detection and false alarm rates during winter 2016–2017 produced an average lead-time predictability of ~6–7 days for various P-O-T values.