Establishment and Comparison of Two Types of Statistical Monthly Spatial Prediction Models for Weddell Sea Ice

Abstract In recent years, the Weddell Sea ice loss accounted for nearly half of the Antarctic sea ice loss, whose prediction has attracted much attention. In this study, we developed two types of monthly spatial statistical prediction models for sea ice concentration (SIC) in the Weddell Sea, specifically the multiple linear regression (MLR) and multivariate empirical orthogonal function (MEOF) prediction models. Both prediction models shared a common set of advanced oceanic and atmospheric variables, including the Pacific decadal oscillation, eastern tropical Pacific Ocean sea surface temperatures (SSTs), western tropical Indian Ocean SST, southern tropical Atlantic Ocean SST, Antarctica sea level pressure, and Amundsen–Weddell dipole of surface air temperatures with the leading times of 1–57 month(s). The first-month Weddell SIC ahead of the predicting time was also individually incorporated or together with the 12th month advanced one as the predicting factors. Although the MEOF method is excellent at extracting the spatial mode of meteorological variables, it is unexpected that the MLR prediction model demonstrated enhanced performances. The median of the average anomaly correlation coefficient (ACC) of the MLR prediction model was up to 0.45 and that of the average RMSE was as low as 0.8, which was significantly better than those of 0.34 and 1.12 of the MEOF prediction model. This study highlights the advantages of the MLR prediction model and the importance of employing advanced sea ice to improve the spatial predictive performance of statistical models.