Author(s): Amelie Bouchat; Pascale Bourbonnais; Philippe Lamontagne; Jean-Francois Lemieux; Bruno Tremblay
Linked Author(s):
Keywords: River Ice; Ice Formation; Growth and Dynamics
Abstract: While the presence of river ice can critically affect navigation safety, forecasts of ice conditions are almost non-existent for the St-Lawrence Seaway, a major shipping channel in Canada. To address this, we investigate sources of predictability for the Seaway freeze-up date (FUD) near Montreal. We identify that predictability originates mostly from atmospheric variables (e.g. air temperature, cloud cover, snowfall, wind speed) and from the North Atlantic Oscillation index, with a maximum lead time of ~2 months. Based on this analysis, we develop statistical models to forecast the FUD at sub-seasonal scales using a Multiple Linear Regression (MLR) approach and a Multi-Layer Perceptron Machine Learning (ML) approach. We also present a categorical FUD forecast using seasonal predictions of air temperature and show that its accuracy is not better than that of a random forecast (< 33%). On the other hand, a naive climatological forecast can already forecast the correct FUD category (i.e. below, near, or above average) with an accuracy of 50% and 5.5 to 6.5 days of mean absolute error (MAE). The ML models are trained to forecast the water temperature, from which the FUD is indirectly detected. We show that this leads to a late bias of ~10 days in the ML forecasts, corresponding to the bias of the freeze-up date derived from the climatological water temperature time series. After correcting for this bias, we show that both ML and MLR can beat the climatological baseline at short lead time (December 1st), but their performance is highly dependent on the evaluation period and, for ML models, on the choice of predictors. We therefore propose to optimize different ML models independently for different lead times and sets of predictors, however we note that the ML approach is likely limited by having too little training data.
Year: 2022