Exposure modelling of transmission towers using street-level imagery and a deep learning object detection model

• Type: Article in International Conference Proceedings Book
• Date: 2022

Authors

Cesarini, L (Author) Other; Figueiredo, R (Author) Other; Xavier Romão (Author) FEUP; Martina, M (Author) Other

Conference proceedings International

Title: Proceedings of the International Conference on Natural Hazards and Infrastructure

3rd International Conference on Natural Hazards and Infrastructure, ICONHIC 2022

5 July 2022 through 7 July 2022

Indexing

Scopus

Other information

Authenticus ID: P-00W-AQM

DOI: 10.5194/egusphere-egu22-9406

Abstract (EN): Exposure modelling is a vital component of disaster risk assessments, providing geospatial information of assets at risk and their characteristics. Detailed information about exposure bring benefits to the spatial representation of a rapidly changing environment and allows decision makers to establish better policies aimed at reducing disaster risk. This work proposes and demonstrates a methodology aimed at linking together 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 of power grid transmission towers, an asset particularly vulnerable to strong wind among other perils. The methodology is implemented through a start-to-end pipeline that starting from the locations of transmission towers derived from the power grid layer of OSM¿s world infrastructure, can assign relevant features of the tower based on the identification and classification returned from an object detection model over street-level imagery of the tower, obtained from GSV. The initial outcomes yielded promising results towards the establishment of the exposure dataset. For the identification task, the YOLOv5 model returned a mean average precision (mAP) of 83.57% at intersection over union (IoU) of 50%. For the classification problem, although predictive performance varies significantly among tower types, we show that high values of mAP can be achieved when there is a sufficiently high number of good quality images with which to train the model. © 2022, National Technical University of Athens. All rights reserved.

Language: English

Type (Professor's evaluation): Scientific

No. of pages: 11