Applying Statistical and Machine Learning Methods for rapid earthquake alert system in Greece with a new mobile application

Confirming the prompt and accurate notification of earthquakes is vital for mitigating their potential impacts. To achieve this, statistical approaches, including Machine Learning (ML), have become indispensable tools across various scientific fields, particularly in Seismology and seismic data. This research explores the utilization of ML techniques to improve earthquake real time alerts. The case study is Greece and the surrounding region, an area with highest seismic activity throughout the Mediterranean.   

This work is focused on the real time collection and processing of an extensive earthquake dataset to generate earthquake alerts by making phone calls and providing details about the time, magnitude, and epicenter of each seismic event. Previous efforts aimed to extend these alerts beyond the notifications (emails and messages) that analysts at the Seismological Center of AUTH (Aristotle University of Thessaloniki) received during their duty. The goal was to make these alerts accessible to all citizens, communities, civil protection agencies and various authorities (e.g. municipalities, schools, police, etc.). The island of Kefalonia served as a pilot region where this functionality was initially implemented. We then chose to extend the application to all Ionian islands to encompass the entire region.

The new insight here is the development of a mobile application that allows users to define a specific geographical region for receiving notifications-alerts. The AI Service will combine the real time earthquake information in conjunction with the geometry defined by each user in order to classify whether a notification should be sent to that specific user.

As training input data used in the application, we first require a catalog of earthquakes spanning from 2011 to 2025 with M≥3.0, along with demographic data for Greece region provided by the Hellenic Statistical Authority. A radius around each epicenter is calculated by considering the earthquake’s macroseismic Intensity (I), the earthquake’s magnitude (M), earthquake depth, total population and number of households within the calculated radius. The labeled dataset is then used to train a classification model via Azure AutoML. This model identifies significant earthquakes and determines which areas to call in order to provide earthquake alert. Notification messages could be to any subscribed mobile number with the calling voice available in Greek, English, or French.