A hybrid ensemble modelling framework for the prediction of breakup ice jams on Northern Canadian Rivers

Abstract

The forecasting of river ice jams faces challenges relating to both the availability of data and the complexity of river ice dynamics, resulting in difficulties in model formulation. In this study, a hybrid ensemble modelling framework is developed to address the data scarcity issue and leverage advanced machine learning techniques for the prediction of ice jams with a one-day lead time. The proposed methodology utilises data easily monitored in advance of any ice jam events and maintains a realistic balance between ice jam and non-ice jam events. A combination of both single model algorithms, including classification trees, logistic regression, k-nearest neighbors, support vector machines, and artificial neural networks, and ensemble model algorithms, including random forest, adaptive boosting, gradient boosting, variance penalizing adaptive boosting, logistic boosting, class switching, and adaptive resampling and combining, are considered for both the member models of the first layer of the hybrid ensemble and for the ensemble combiner of the second layer. The final selection of both variables and member models for the hybrid ensemble is detailed, with a focus on the reduction of false negatives, the prediction of no ice jam on a day when one occurs. The proposed method is applied to the St. John River in New Brunswick, Canada, in a location particularly prone to ice jam flooding. Using the proposed methodology, a final model combining 6 different member models using a support vector machine as the ensemble combiner was produced, with a balanced prediction accuracy of 86%. This hybrid ensemble model outperformed the other tested ensemble models, as well as a series of generalized models produced using all available input variables and member models. The model also performed well against other ensemble techniques and against the individual member models. These results demonstrate the viability of the proposed methodology in constructing a hybrid ensemble model for the forecasting of ice jams on Northern Canadian Rivers. The techniques utilised can be adapted to other locations to facilitate ice jam forecasting, requiring data that is easily available and monitored in advance of any potential flooding events.