A deep learning based prediction of Arctic sea ice concentration using satellite and reanalysis data

Abstract

Arctic sea ice is one of the key factors closely related to climate change and energy balance. The sea ice concentration (SIC) can be used to explore the spatial distribution of sea ice through time. Arctic SIC is highly related to surrounding environments such as atmosphere, ocean, and climate change. Thus, the prediction of Arctic SIC can provide the crucial information related to the dynamic environment of the Arctic and Earth systems. Previous studies have suggested several approaches such as physical and statistical models to predict Arctic SIC. Due to the high complexity of Arctic environments, it is hard to figure out the interaction between sea ice and the atmosphere and/or ocean in detail. Thus, this study suggests a data-driven model using machine learning approaches for predicting SIC. Convolutional neural networks (CNN) and random forest (RF) were used to predict monthly SIC with multiple satellite products and reanalysis data. During the melting season (Jun. to Sep.) of past 16 years (2002-2017), monthly mean data of SIC (NOAA OISST v2), ice temperature (MODIS), and other atmospheric and oceanic factors (ECMWF ERA-interim reanalysis data) were used to predict SIC of the next month. Both models were built with an extensive dataset of Arctic SIC and evaluated using the cross-validation by year. Over the melting season, the performance of CNN was better than RF (4.7% and 7.5% of RMSE, respectively). In the literature, a challenging problem exists when predicting SIC around the ice edge due to its high variability over time. Around the ice edge (latitude < 80°), the CNN model showed a significantly improved result than RF.