Machine learning for prompt estimation of macroseismic intensity from seismometric data in Italy

TL;DR

This paper introduces a machine learning method to quickly estimate earthquake intensity in Italy using seismic data, improving response and damage assessment.

Contribution

A novel Random Forest framework with surrogate decision trees for interpretable macroseismic intensity estimation from early seismic data.

Findings

The Random Forest model outperformed existing methods in predicting macroseismic intensity.

Surrogate decision trees provided interpretable insights into the model's predictive mechanisms.

The model's uncertainty was assessed, enhancing reliability for real-time decision-making.

Abstract

After an earthquake, it is crucial to rapidly and accurately estimate macroseismic intensity to guide rescue operations and assess potential damage. The Mercalli-Cancani-Sieberg intensity scale is used to qualitatively assess the ground shaking based on observed effects. This study develops a Machine Learning framework, leveraging the Random Forest algorithm, to estimate macroseismic intensity using early available seismic data. Data from different sources are used for model training: seismic data from the Italian instrumental monitoring networks of Istituto Nazionale di Geofisica e Vulcanologia and Protezione Civile, as well as macroseismic intensity data from both the online macroseismic questionnaire and the on-site surveys by field experts. In order to explain the predictive mechanism of the Random Forest algorithm, this study makes use of surrogate decision trees, providing an interpretative key for the informed decision-making process during seismic events. These models provide insights into the relationships between covariates and predicted intensities, enabling the discussion of model complexity, predictive capability, and explainability. Furthermore, the uncertainty associated with the predictions of the surrogate trees is assessed. When compared with other models for estimating intensity based on ground motion peaks or source parameters, the Random Forest model achieved better predictive performance.

The online version contains supplementary material available at 10.1038/s41598-026-35740-x.

Figures