Enhancing ocean monitoring for coastal communities using AI

Abstract

This research addresses the challenge of estimating significant wave height (SWH) using seismic data; a task beneficial for coastal safety, maritime operations, and environmental monitoring. Given that the livelihoods of over three billion people depend on coastal and marine resources, accurate and accessible models for SWH prediction are essential for informed decision-making and disaster preparedness. This study revisits and improves upon existing methods for the prediction of SWH using seismic data, specifically focusing on the limitations of data quality and computational feasibility in resource-constrained settings. The problem was tackled by first formulating a baseline method, followed by the identification and implementation of key innovations. These included the use of a longest-stretch algorithm for more accurate seismic data, and hyperparameter tuning tailored to the local characteristics of each seismic station. The seven final models were trained and evaluated on consumer-grade hardware, ensuring their accessibility for deployment in areas with limited resources. These models were rigorously evaluated against existing baselines, with performance metrics including the coefficient of determination (R²) and mean absolute error (MAE). The seven final models achieved a mean R² of 0.82978 (minimum: 0.60686, maximum: 0.92060) and a mean MAE of 0.13476 (minimum: 0.10066, maximum: 0.18243). The results demonstrate significant improvements over the baseline models, particularly in terms of predictive accuracy and model efficiency. Specifically, the final models achieved an increase in R² of up to 0.13316 and a reduction in MAE by 0.02625 m over the baseline models, considering the average performance across all stations. However, challenges such as accurately predicting extreme weather still remain, due to the limited existence of such instances within the data. The primary contribution of this research is the development of computationally efficient, locally optimised models for SWH prediction using seismic data, covering a region of interest around Sicily and Malta, that can be deployed on consumer-grade hardware. This expands the accessibility of artificial intelligence in low-resource settings. Future work could focus on advanced gap-filling techniques, data augmentation for extreme weather scenarios, and alternative model architectures for better handling extreme values.