Integration of LiDAR and multispectral images for exposure and earthquake vulnerability estimation. Application in Lorca, Spain

TL;DR

This paper introduces a scalable method combining LiDAR and multispectral imagery with machine learning to assess urban seismic vulnerability, validated on Lorca, Spain, achieving around 80% accuracy.

Contribution

It presents a novel integrated procedure that combines image segmentation, building attribute extraction, and machine learning for seismic vulnerability estimation.

Findings

Machine learning techniques achieved 77-80% classification accuracy.

The method is scalable and adaptable to different urban areas.

It provides a cost-effective alternative to traditional field surveys.

Abstract

We present a procedure for assessing the urban exposure and seismic vulnerability that integrates LiDAR data with aerial and satellite images. It comprises three phases: first, we segment the satellite image to divide the study area into different urban patterns. Second, we extract building footprints and attributes that represent the type of building of each urban pattern. Finally, we assign the seismic vulnerability to each building using different machine-learning techniques: Decision trees, SVM, logistic regression and Bayesian networks. We apply the procedure to 826 buildings in the city of Lorca (SE Spain), where we count on a vulnerability database that we use as ground truth for the validation of results. The outcomes show that the machine learning techniques have similar performance, yielding vulnerability classification results with an accuracy of 77% - 80% (F1-Score). The procedure is scalable and can be replicated in different areas. It is especially interesting as a complement to conventional data gathering approaches for disaster risk applications in areas where field surveys need to be restricted to certain areas, dates or budget. Keywords LiDAR, satellite image, orthophoto, image segmentation, machine learning, earthquake vulnerability.