Ice-jam flood predictions using an interpretable machine learning approach

Machine-learning algorithms have been employed in river ice research for flood estimation. This study aimed to introduce a machine learning-based model for predicting ice jam floods. An ice-jam dataset was created using a stochastic modelling approach in which thousands of possible scenarios were simulated. This approach integrated a hydrodynamic model, RIVICE, into a Monte Carlo Analysis (MOCA) framework. The set of parameters, boundary conditions, and associated backwater level elevations was then applied to a machine-learning algorithm to implement a preliminary ice-jam flood prediction model, combining decision tree regressors (DTR) with an adaptive boosting (AdaBoost) regressor. Shapley Additive explanations (SHAP) were then applied in the preliminary model to identify the most influential parameters of ice-jam flooding. Identified variables from SHAP were then used to construct a simple ice-jam flood hazard prediction model with fewer variables. The Athabasca River in Fort McMurray, Canada, is a test site for this modelling framework.