An Improved ConvLSTM Network for Arctic Sea Ice Concentration Prediction

An accurate prediction of sea ice concentration (SIC) is highly necessary to establish the Arctic shipping routes, polar research, and scientific operations effectively. Presently, the high nonlinearity, complexity, and discontinuity of polar SIC, have contributed to the poor timeliness and inaccurate results in SIC prediction. Traditional SIC prediction methodology mostly focuses on its temporal evolution rule but neglects the analysis of spatial information for forecasting. In this paper, we develop an improved convolutional long short-term memory (ConvLSTM) network with multi-layer stacking, an effective spatiotemporal forecast network, for short- and mid-term prediction in Arctic SIC. This network is capable of processing spatial information utilizing the convolution operation in ConvLSTM to obtain the SIC spatial feature relations of the sea domain. The result indicated that the correlation coefficient (r) of the 7-day prediction of the 3-layer stacking architecture ConvLSTM network is as high as 97.57%. The model proves to have good performance in spatiotemporal prediction, which provides new insights in the SIC prediction field.