Forecasting Antarctic Sea Ice Concentrations Using Results of Temporal Mixture Analysis

Abstract

Sea ice concentration (SIC) data acquired by passive microwave sensors at daily temporal frequencies over extended areas provide seasonal characteristics of sea ice dynamics and play a key role as an indicator of global climate trends; however, it is typically challenging to study long-term time series. Of the various advanced remote sensing techniques that address this issue, temporal mixture analysis (TMA) methods are often used to investigate the temporal characteristics of environmental factors, including SICs in the case of the present study. This study aims to forecast daily SICs for one year using a combination of TMA and time series modeling in two stages. First, we identify temporally meaningful sea ice signatures, referred to as temporal endmembers, using machine learning algorithms, and then we decompose each pixel into a linear combination of temporal endmembers. Using these corresponding fractional abundances of endmembers, we apply a autoregressive model that generally fits all Antarctic SIC data for 1979 to 2013 to forecast SIC values for 2014. We compare our results using the proposed approach based on daily SIC data reconstructed from real fractional abundances derived from a pixel unmixing method and temporal endmember signatures. The proposed method successfully forecasts new fractional abundance values, and the resulting images are qualitatively and quantitatively similar to the reference data.