This perspective examines the application of transfer learning (TL) within deep learning (DL) frameworks for extreme water level (EWL) and spatiotemporal flood predictions. We discuss the main advantages of TL, such as enabling model transferability of pretrained DL models from/to diverse coastal-estuarine systems and reducing computational time compared to physics-based models. These advantages can accelerate the deployment of flood prediction models in data-limited locations. We also discuss challenges and limitations that hinder accurate pattern recognition and propagation of EWLs from gauge (observation) stations to surrounding locations within model domains. These limitations include dependence on similarity in data distributions, overfitting the training data and both lag and hysteresis effects in the timing of peak water levels and flood dynamics. Lastly, we explain several misconceptions and challenges in current DL approaches that hinder EWL and spatiotemporal flood prediction, including training models exclusively on extreme conditions, assessing accuracy solely through goodness-of-fit metrics, and connecting the model’s knowledge of input data with physical explanations of flood processes without an adequate context. We argue that these challenges can be addressed by prioritizing storm-relevant patterns of the input data features as well as embedding spatial propagation of EWLs in DL frameworks to mimic coastal-estuarine hydrodynamic models. Ultimately, progress toward generalizable model transferability relies on the modeler’s ability to incorporate physical understanding in DL architectures, alongside continued advances in physics-informed machine learning models via soft or hard constraint approaches. There remains substantial work to establish guidelines and/or formal procedures to develop robust, interpretable, and generalizable DL models for spatiotemporal flood prediction, thereby supporting effective flood management, mitigation, and emergency preparedness.

We analyse a 36-year hydrodynamic and morphological dataset from the Hasaki coast, Japan, comprising 501 wave storm events (405 individual and 96 clustered events) to investigate the impact of storm dynamics and clustering on beach erosion. Focusing on the wave component of storms, events are identified using wave height thresholds. Daily and weekly beach profile measurements from the Hasaki Oceanographic Research Station are used to quantify erosion. The study examines the seasonal influences on Hasaki beach, the characteristics and temporal evolution of storms, and their associated erosional impacts. Moreover, we test two supervised machine learning (ML) algorithms, support vector regression (SVR), and deep neural network (DNN), in predicting shoreline change using 16 wave, storm, and morphological features. SVR showed reasonable accuracy on the training dataset but underperformed on testing, while DNN failed to produce reliable predictions on both. With SVR yielding an R2 of 0.18 and DNN 0.27 on the testing dataset, we conclude that, given the limited data and available features, such ML models may not generalise well. However, separate analyses using observed data reveal clear seasonal variations in wave storm dynamics and distinct behaviours of clustered events associated with beach erosion, highlighting important insights beyond the ML results.

Low-lying atoll islands are among the world’s most vulnerable coastal environments to sea-level rise (SLR). Global application of coastal flooding models suggests that centennial flood events may become annual events by 2050 in tropical regions. This article addresses this claim by modelling an island flooding event that occurred in the Maldives on 1 July 2022 as a result of a distant-swell event coinciding with an extra high spring tide. Hydrodynamic data collected after the event on one of the affected islands were used to calibrate and validate a one-dimensional non-hydrostatic XBeach model. The model overpredicted wave setup and underpredicted the water motion at frequencies <0.05 Hz, but the wave run-up elevation was predicted reasonably well. The 1 July flood event was considered in a decadal context using modelled wave data and measured tide data. It was concluded that the 1 July event represents a c. 1:25-year flooding event, but, due to SLR, such flooding could occur every few years by 2050. This prediction ignores natural or anthropogenic adjustments to the island morphology. The expected increase in frequency of coastal flooding in the Maldives requires atoll and island authorities in the Maldives to act swiftly in adapting to future flood risk.

The June 2016 extratropical cyclone with anomalous ENE storm wave direction caused widespread beach-foredune erosion in southeastern Australia. At Bengello Beach, erosion volumes were 97–108 m3/m for the central and southern parts of the beach, while the northern end only lost 18 m3/m of sand. In the central and southern parts of the embayment, a surf zone bar formed 50–100 m further seaward than is typical for this beach and was a temporary store of sand eroded from the beach-foredune. A nearshore wave model showed substantial variability in wave power along the 10 m depth contour and explained the partial sheltering of the northern end of the embayment from storm impact. An embayment-wide time-series of airborne LiDAR further emphasised the alongshore variability in beach-foredune erosion. The wide beach and broad, double-crested, well-vegetated foredune along much of the embayment was pivotal in protecting the shoreline. In the centre and south of the beach, recovery took nearly three years and although complete by volume, the foredune was narrower and less resilient. The results emphasise the role of wide beaches and natural vegetated foredunes in buffering extreme storms and suggest foredune rehabilitation should be a key management priority for sustainable coasts.