Predicting Sea Ice Conditions using Neural Networks

Abstract

Safe and efficient marine navigation in ice-infested waters requires comprehensive and timely information on the sea ice conditions. These include information on the ice concentration and type, ice edge location, icebergs and open leads. The Canadian Ice Service is responsible for providing ice information in Canadian waters, mainly through its daily ice charts. Unfortunately, due to the difference in time between the ice chart production and its use by mariners, the ice information is always out of date. This problem might be overcome by developing a neural network-based system for predicting the ice conditions over time. A supervised neural network is trained to predict the ice conditions at a given location and time using the current ice charts, which are provided by the Canadian Ice Service. The input ice data is mapped to an output vector that gives the predicted ice conditions. The traditional non-modular feed-forward neural network structure failed to map the required function, and hence, was modularized to give better prediction performance. Each neural module was responsible for the prediction of a 5×5 km area, while the ice characteristic of interest was the total concentration.