Application of a context augmented autoencoder to Arctic sea ice forecasting

We study the importance of using modern techniques for training deep neural networks on an example of the task of forecasting sea ice concentration in the Arctic region of the Barents and Kara Seas. Focusing on a context-augmented autoencoder, we show the importance of disentangling the model’s latent representations in this task. In addition, we show the benefits of using a pre-trained network checkpoint by demonstrating faster convergence of the training process to a more optimal extremum. As a practical result, we obtain a data-driven sea ice forecasting system based on new architectural principles - an LSTM network trained by autoencoder latent representations to predict 2-dimensional sea ice maps.