A Bayesian Categorical Regression Model for Probabilistic Predictions of Minimum Sea Ice

Recent decline of Arctic sea ice extent has increased the prospect of an ice free arctic. The retreat of sea ice offers opportunities (shorter shipping lanes, tourism, etc.) and geopolitical challenges among the Arctic Nations and near Arctic observers. Planning for these opportunities and challenges requires understanding of the space-time variability of sea ice attributes along with skillful predictability, especially of the September minimum sea ice extent. Presently, dynamic models are the primary tools for this endeavor with varying degrees of skill. Here we introduce a statistical model using Generalized Linear Model (GLM) framework with Bayesian formulation that is capable of providing skillful probabilistic forecasts of sea ice cover, along with quantifying the attendant uncertainties. The presence or absence of ice (absence defined as ice concentration below 15%) is modeled using a categorical regression model, which is a subset of the GLM framework with local covariates (surface air temperature, geopotential height at 500hpa, sea ice concentration) at varying lead times, along with time index to capture the temporal trend. The model parameters are estimated in a Bayesian formulation thus enabling the posterior predictive probabilities of the sea ice extent. The model is fitted and validated to September minimum sea ice concentration data from 1980 through 2017. This fitted model is used to predict the minimum September sea ice cover in 2018. The model demonstrates skillful predictions at 3-month lead-time, which is of immense use to various sectors - defense, tourism, shipping etc. The Bayesian framework also provides a novel approach to understand and model the space-time variability of Arctic sea ice.