Calibration of subseasonal sea‐ice forecasts using ensemble model output statistics and observational uncertainty

Abstract

In response to a growing demand for improved sea‐ice forecast guidance at shorter timescales and higher spatial resolutions, this study investigates the predictive skill of daily subseasonal sea‐ice forecasts from two state‐of‐the‐art prediction systems: SEAS5 from the European Centre for Medium‐Range Weather Forecasts (ECMWF) and the Global Ensemble Prediction System (GEPS) of Environment and Climate Change Canada (ECCC). Based on hindcast records from 1998–2017, we find that probabilistic forecasts of sea‐ice concentration (SIC) throughout the marginal ice zone of the Canadian Arctic in June and November are no more skillful than simple benchmark forecasts based on climatology and damped persistence. At short lead times, the lack of skill arises from overconfident ensemble spread and errors in forecast initial conditions. At longer lead times, the development of model drift also plays a role. To improve the forecasts, we develop the nonhomogeneous censored Gaussian regression for SIC (NCGR‐sic) calibration procedure, which uses the forecast ensemble mean and ensemble standard deviation (both locally and from neighboring locations) as predictors in an ensemble model output statistics framework. Importantly, NCGR‐sic incorporates observational uncertainty directly during model parameter fitting, leading to enhanced calibration. NCGR‐sic improves the spatial probability score by ∼60%$$ \sim 60\% $$, on average, half of which we infer results from the removal of climatological bias and half from the improvement of forecast uncertainty. Calibrated forecasts from SEAS5 (GEPS) exceed the skill of climatology throughout the marginal ice zone over at least the first 33 (26) days in June and the first 32 (15) days in November. Generally, better skill occurs outside the Canadian Archipelago and for low‐ to mid‐range SIC compared with high SIC. While calibrated forecasts also outperform damped persistence after the first 3–10 days, we postulate that initial condition inconsistencies must be resolved in order to obtain skill at shorter timescales. Finally, the robustness of NCGR‐sic is demonstrated by effectively improving an operational forecast from June 2020.

Deep Learning‐Based Seasonal Forecast of Sea Ice Considering Atmospheric Conditions

“…Information about SEAS5, Météo France, and GloSea6 is detailed in Table S1 in Supporting Information . The dynamical seasonal forecasting systems show relatively low forecast accuracy, which may be due to uncertainties in model physics (Blanchard‐Wrigglesworth et al., 2015), initial conditions (Blanchard‐Wrigglesworth et al., 2017), and model drift (Dirkson et al., 2022). By comparison, Atsicn shows a significant improvement in SIC forecast, indicating the advantages of data‐driven modeling in effectively learning the intricate and nonlinear spatiotemporal patterns of the seasonal sea‐ice system.…”

Understanding sources of Northern Hemisphere uncertainty and forecast error in a medium‐range coupled ensemble sea‐ice prediction system

“…Bias-correction methods can be challenging in the context of a rapidly evolving sea-ice background state due to climate change, where typically skill improvement is reliant upon improving climatological bias, with the underlying assumption that previous (hindcast) bias is representative of forecast bias. Dirkson et al (2022) has demonstrated that, at least for more robust calibration methods, forecast skill can still be improved. Still, accurate assessments of uncalibrated forecast skill are necessary to assess the system for operational purposes.…”

“…Dirkson et al . (2022) has demonstrated that, at least for more robust calibration methods, forecast skill can still be improved. Still, accurate assessments of uncalibrated forecast skill are necessary to assess the system for operational purposes.…”

Calibration of short-term sea ice concentration forecasts using deep learning

“…In sea ice forecasting, most calibration methods have been developed for subseasonal to seasonal time scales (e.g. Zhao et al, 2020;Director et al, 2021;Dirkson et al, 2019Dirkson et al, , 2022, but short-term sea ice forecasts produced by dynamical models are usually not calibrated despite their potential interests for end-users (Wagner et al, 2020). Nevertheless, Palerme and Müller (2021) showed that the errors of short-term sea ice drift forecasts (up to 10 days) from the TOPAZ4 prediction system (Sakov et al, 2012) can be significantly reduced using random forest models (by 8 % and 7 % for the direction and speed of sea ice drift, respectively).…”