Blind testing of shoreline evolution models.

Area of Science:

• Coastal geomorphology
• Computational modeling
• Machine learning applications

Background:

• Beaches globally experience dynamic changes due to waves and tides.
• Accurate prediction of shoreline evolution is difficult, even for short-term (decadal) forecasts.

Purpose of the Study:

• To evaluate the predictive capabilities of various numerical and machine learning models for shoreline evolution.
• To compare model performance using real-world data from Tairua beach, New Zealand.

Main Methods:

• A modeling competition involving 19 diverse numerical models and machine learning techniques.
• Utilized 18 years of daily shoreline position and beach rotation data from a camera system.
• Models were calibrated on historical data (1999-2014) and tested on unseen forecast data (2014-2017).

Main Results:

• Both traditional and machine learning models accurately reproduced shoreline changes during normal conditions in the calibration period.
• Predictive accuracy decreased for both approaches during the forecast period, particularly for some machine learning algorithms.
• Model ensembles demonstrated superior performance and allowed for uncertainty assessment.

Conclusions:

• Current shoreline evolution models face limitations in predicting extreme and rapid coastal changes.
• Model ensembles offer a more robust approach to forecasting and uncertainty quantification.
• Collaborative research initiatives like modeling competitions are crucial for advancing coastal prediction science.