Prediction of sea ice concentration using artificial neural networks

Artificial neural networks are computational models capable of solving complex problems through learning, or training, and then generalizing the network solution for other inputs. This thesis examines the performance of two neural network-based models, which were developed for predicting the ice concentration in the Gulf of St. Lawrence in Eastern Canada. The first is a batch model which uses time to predict future ice concentration, while the second model predicts the ice concentration sequentially. It is shown that the performance of the two models is almost identical, as long as no abrupt changes occur in the ice conditions. If, however, the ice condition changes suddenly, only the sequential model is proved to be capable of predicting the ice condition without noticeable accuracy degradation. A performance comparison is made between the developed neural network model and coupled ice-ocean model for ice concentration prediction to further validate the model.