Morphodynamic predictions based on Machine Learning. Performance and limits for pocket beaches near the Bilbao port

Abstract

Predicting the morphodynamic behaviour of pocket beaches exposed to energetic waves and meso-tidal ranges-particularly under strong seasonal variability and the influence of climate change-requires a robust characterization of coastal morphodynamics across a wide range of temporal and spatial scales. This study introduces a data-driven modelling approach using Machine Learning (ML), specifically the Gradient Boosting Regressor (GBR), a powerful ensemble technique capable of iteratively improving predictions from limited datasets. The GBR model is applied to forecast beach evolution in complex coastal settings, where physical understanding is limited, specifically targeting a set of pocket beaches in the Bay of Biscay (North Atlantic). The methodology combines wave time series and morphodynamic variables obtained through videometry stations (KOSTASystem technology). This ML framework is then implemented to improve the current understanding of hydro-morphological interactions and establish criteria to enhance the reliability of erosion and flood predictions. The obtained predictions can steer the design and implementation of protection measures to increase beach resilience under climate change drivers, such as sea-level rise and wave storminess, leading to improved adaptation strategies. This approach, which also demonstrates the advantages of ML over conventional statistics, is developed from a set of extreme meteo-oceanographic events acting on pocket beaches adjacent to and within the Nervi \& oacute;n estuary and Bilbao port. The application of conventional statistics and ML techniques to this dataset begins with an extreme analysis of offshore wave data, from which a set of 32 wave storms has been propagated towards the coast using the Simulated WAves Nearshore (SWAN) model. This dataset serves to evaluate predictive formulations derived from statistical and ML tools, based on monthly values, which filter out short-term variability and focus on medium- to long-term (annual to decadal) beach behaviour-scales that are critical for sustainable coastal management. Results demonstrate that ML-based predictions using GBR outperform traditional statistical methods, where validation metrics confirm the improved predictive accuracy, with R2 values exceeding 0.7 in several cases, without any evidence of overfitting. These predictions contribute to understanding hydro-morphological interactions and support the design of adaptive beach protection strategies.