Machine Learning for Meteotsunami Prediction: A Case Study on the Portuguese Coast

Meteotsunamis, tsunami-like waves triggered by rapid atmospheric pressure disturbances, can result in significant coastal damages. This study introduces a machine learning (ML) framework for predicting meteotsunami occurrences along the Portuguese coast, using both atmospheric pressure records and tide gauge data collected from 2010 to 2020. A methodology is proposed to construct a structured dataset of inputs and targets from continuous meteorological and sea-level observations, yielding an imbalanced dataset with a meteotsunami-to-nonevent ratio of approximately 1:60. To address this imbalance, class weighting and an ensemble strategy aggregating predictions across multiple observatories were implemented in the ML framework. 

The prediction model employs an encoder-decoder architecture, integrating Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) layers. Results demonstrate the model's effectiveness in  capturing the complex dynamics of meteotsunami formation and propagation  with accuracy and reliability for operational forecasting. Future research will focus on incorporating additional meteorological variables such as wind speed and direction, expanding the spatial and temporal coverage of data, and further refining prediction capabilities to enhance meteotsunami early warning systems and mitigate meteotsunami-related risks.