Data-driven proactive prediction of pumice drifting patterns using similarity search of the Kuroshio current axis

Abstract

Abstract Pumice drifting poses substantial risks to maritime navigation and coastal communities. While traditional ocean-current-based simulations effectively predict drifting patterns, they are resource-intensive and unsuitable for real-time use following abrupt eruptions. This study proposes a data-driven framework that enables rapid, low-cost pumice drift prediction, leveraging daily-reported Kuroshio Current axis (KCA) patterns and conducting similarity searches on pre-existing simulation datasets. Focusing on eruptions at Bayonnaise Rocks in the Izu Islands, Japan, we demonstrate that the dynamic time warping distance, a measure of similarity between the current KCA pattern and historical KCA patterns, accurately predicts drifting scenarios within the critical first 10 days post-eruption. This method reliably identifies high-risk cases, including those affecting coastal regions, without requiring new simulations. By refining simulation datasets and enhancing prediction accuracy, this framework can become a practical tool for hazard assessments, offering a scalable solution for proactive disaster-risk management in response to unpredictable pumice eruptions.