Abstract
Marine operations in ice-covered waters require reliable and timely information about the sea ice conditions. The Canadian Ice Service produces and distributes the ice information to mariners operating in the Canadian water in the form of daily ice charts. Unfortunately, however, due to the time difference between the production and the use of the ice charts, the ice information is always out of date, which endangers the safety of marine operations. To efficiently overcome this problem, a reliable model for predicting the sea ice conditions (concentrations) over time is developed. Inspecting the ice charts for the period 1987 to 1998 showed that the sea ice conditions change according to a regular pattern to some extent. Therefore, a neutral network function approximation system could model, and hence predict, these changes efficiently when trained using multiple-year ice concentrations readings. The data used in training the neural network are extracted from the ice charts for the Gulf of St. Lawrence in eastern Canada. The input to the network is a vector which represents the current ice concentrations over a test area containing 40 points. The input vector is mapped to an output vector that gives the predicted ice concentrations.