An intelligent prediction model for bank retreat based on hydro-sediment dynamics mechanism and machine learning

Severe riverbank retreat poses a significant threat to channel stability and flood control safety. To address the challenge of quantitatively predicting of bank retreat, an intelligent prediction framework was developed integrating physical mechanisms with machine learning. The framework enables quantitative forecasts of both the probability and the width of bank retreat. To overcome the limited feature-representation capability of conventional machine learning algorithms, a multidimensional feature system was established based on the dynamics of bank failure, incorporating flow and sediment conditions, channel boundary conditions, and preceding morphological changes of the riverbed. By coupling the K-nearest neighbor (KNN) algorithm and Lasso regression, two models were constructed: a bank-retreat probability model and a retreat-width model, both equipped with an embedded feature-selection module. The models were applied to a braided reach of the Lower Yellow River between Tiexie and Gaocun, using observed hydrological and cross-sectional topographic data from 1999 to 2024. Retreat widths were categorized into mild (＜50 m), strong (50-110 m), and severe (＞110 m) according to their probability distribution. The results indicate that the models can reliably predict both the occurrence and magnitude of bank retreat. The probability-prediction model achieved accuracies of 0.97, 0.80, and 0.86 for the training, validation, and test sets, respectively. The width-prediction model yielded mean absolute errors of 13 m, 48 m, and 48 m for the corresponding sets, and obtained an accuracy of 0.83 for retreat-level classification in the test set. Applied to the 2024 flood season, the models successfully predicted 70% of the bank-retreat events, with an accuracy of 0.71 for retreat-severity classification. The intelligent prediction framework developed in this study effectively addresses the challenge of quantitatively predicting bank retreat. It demonstrates high accuracy and good applicability in predicting bank-retreat processes in alluvial rivers, providing valuable technical support for ensuring channel stability and flood-control safety.