Mid-Term Seasonal Arctic Sea Ice Concentration Forecasting Based on CNN-ConvLSTM and Wavelet Multi-Scale Deep Learning Algorithms

Mid-term seasonal Arctic sea ice concentration forecasts are vital for navigation routing, policy-making, and predicting cetacean migrations for safety reasons. The accuracy of such predictions typically declines with extended lead times. Notably, forecasting during the melting season presents more challenges than the freezing season, due to sea ice thinning. In this research, we try to address these challenges by employing convolutional neural networks (CNN) and convolutional long short-term memory networks (ConvLSTM) in M-band wavelet domain and integrating three-dimensional spatio-temporal climate and sea ice data for enhancing mid-term sea ice concentration (SIC) forecast. Our experimental results revealed that ConvLSTM consistently surpassed CNN. Moreover, using classification rather thanregression models, the inclusion of climate variables, expanded historical data, and pre-processing of our spatial data with
$M$
-band wavelet transform further enhanced model efficacy.