Forecasting Arctic Sea Ice Concentration using Long Short-term Memory Networks

Due to global warming, Arctic sea ice is now declining, and this loss is a self-accelerating process that speeds up sea ice melting and the severity of climate change. Accurate and timely sea ice information is critically important for better monitoring of global climate. Publicly available multi-source, multi-scale, and high-dimensional sea ice data from satellites is a game changer that allows researchers to better understand the Arctic through more sophisticated methods. This study proposes two Long short-term memory (LSTM) networks for sea ice concentration (SIC) forecasting in the arctic area over 1-, 3-, 6-, and 9-month forecast horizons. The first network forecasts the SIC of each grid in a single output with the grid coordinate must be supplied as an additional input, while the second network forecasts the SIC of all grids at once in a single output. The models with and without atmospheric and oceanic variables as external predictors were trained by using 43 years of data and tuned by using random search strategies. The model performance was evaluated and compared based on the root mean square errors and weighted absolute percentage errors to determine the impact of using climate variables in the prediction and arrive at the best-performing forecast model.