How Do decision-tree-based Machine Learning Techniques Compare To Hybrid Approaches for Predicting Fluvial Dike Breach Discharge??

The breaching of a fluvial dike can have devastating consequences for flooded areas. Accurate prediction of the breach discharge is crucial for enhancing preventive measures and emergency planning. So far, most studies have relied on empirical formulas developed for simplified configurations, which fail to capture the complexity of a real dike breaching event. In this context, machine learning (ML) models offer promising predictive capabilities. This study focuses on decision-tree-based models, trained on 43 dike breaching laboratory tests, and compares three predictive approaches: (1) direct prediction using ML, (2) direct prediction using empirical formulas developed for simplified configurations, and (3) a novel analytical approach with an empirical parameter computed using ML or empirical regressions. The extremely randomized trees algorithm demonstrates particularly high accuracy when predicting the breach discharge (approach 1), while the empirical formulas (approach 2) perform poorly. The new analytical model (approach 3) provides intermediate accuracy. Additionally, a novel hybrid approach is proposed, which consists in applying a ML-based corrective term to approaches (2) and (3). This strategy significantly improves the accuracy of the results associated with the empirical formulas (approach 2) and the analytical model (approach 3), both in interpolation and extrapolation, i.e., when tested inside or outside the ML training space. This makes it particularly valuable for predicting the breach discharge beyond the training space, where ML techniques alone are expected to be less effective. The definition of the dike breach invert level, i.e., one of the model inputs, was varied, but it had little influence on the models’ performance. Expanding the experimental dataset by conducting new laboratory or field tests would further enhance the accuracy and reliability of the ML models. Future studies may explore alternative ML models, including physics-guided deep learning algorithms, which, although in their early stages, hold substantial potential for future applications in predicting the breach discharge outside the model training space.