Exposure Modeling of Transmission Towers for Large-Scale Natural Hazard Risk Assessments Based on Deep-Learning Object Detection Models

Abstract

Exposure modeling plays a crucial role in disaster risk assessments by providing geospatial information about assets at risk and their characteristics. Detailed exposure data enhances the spatial representation of a rapidly changing environment, enabling decision-makers to develop effective policies for reducing disaster risk. This work proposes and demonstrates a methodology linking volunteered geographic information from OpenStreetMap (OSM), street-level imagery from Google Street View (GSV), and deep learning object detection models into the automated creation of exposure datasets for power grid transmission towers, assets particularly vulnerable to strong wind, and other perils. Specifically, the methodology is implemented through a start-to-end pipeline that starts from the locations of transmission towers derived from OSM data to obtain GSV images capturing the towers in a given region, based on which their relevant features for risk assessment purposes are determined using two families of object detection models, i.e., single-stage and two-stage detectors. Both models adopted herein, You Only Look Once version 5 (YOLOv5) and Detectron2, achieved high values of mean average precision (mAP) for the identification task (83.67% and 88.64%, respectively), while Detectron2 was found to outperform YOLOv5 for the classification task with a mAP of 64.89% against a 50.62% of the single-stage detector. When applied to a pilot study area in northern Portugal comprising approximately 5.800 towers, the two-stage detector also exhibited higher confidence in its detection on a larger part of the study area, highlighting the potential of the approach for large-scale exposure modeling of transmission towers.