Applicability of multilayer feed-forward neural networks to model the onset of river breakup

Abstract

Predicting the onset of breakup is an essential component of any ice jam flood forecasting system, yet it presents a difficult challenge due to the complex nature of the relationship between meteorological conditions, streamflow hydraulics and ice mechanics. For this research, data extracted from historical hydrometric and meteorological records were used to develop and assess a three-layer feed-forward artificial neural network (ANN) model for predicting the onset of breakup, using the Hay River in northern Canada as the demonstration site. The calibration results illustrate the potential of the ANN model for successful forecasting of the onset of river ice breakup, i.e. the first transverse cracking of the ice cover. However, rigorous validation also indicates that the accuracy of such ANN models can be optimistically overestimated by their performance during the calibration phase. The possible reasons for this poor predictive capability of the ANN model are also discussed. Despite this caveat, the proposed model shows improved performance as compared to the more conventional multiple linear regression (MLR) techniques typically applied to this problem. ► Illustrates a feed-forward ANN model for forecasting the onset of river breakup. ► Model performance is demonstrated for the Hay River, NWT, Canada. ► ANN model out-performs traditional multiple linear regression models. ► ANN model is validated using the leave-one-out-cross-validation method.