Citation

Zhai W, Peng ZR. Appl. Geogr. 2020; 123: e102252.

Copyright

(Copyright © 2020, Elsevier Publishing)

DOI

10.1016/j.apgeog.2020.102252

PMID

unavailable

Abstract

Assessing damage on the ground is a challenging task for humanitarian organizations and disaster managers due to the limited availability of data and methods for processing. As the most commonly adopted data source, remote sensing imagery can only reflect the damage situation on top of a building and fails to present the damage level from the perspective of the human eye. Recently, an increasing number of Google Street View (GSV) images provide the chance to understand the human's perception of damage on the ground. However, to automatically and quantitatively apply GSV images in damage assessment, two research questions need to be answered: (1) Can deep learning be successfully applied to automate the process of evaluating postdisaster damage using GSV images? (2) Does damage assessment using GSV images provide a different insight, compared with existing approaches, such as remote sensing imagery? Based on our experiments using GSV images and remote sensing imagery in Mexico Beach, FL after Hurricane Michael, we present two conclusions: (1) By applying a deep learning model, the GSV-based damage assessment can be satisfactorily and automatically conducted, with an accuracy of approximately 70% for a single GSV image. (2) GSV images provide a different insight into damage assessment since remote sensing imagery cannot record the damage to exterior walls, windows, doors and facades. When the overall damage level is relatively low, GSV images show better performance in damage assessment. Conversely, when the overall damage level is relatively high, remote sensing imagery shows better performance based our experiments.

Language: en

Keywords

Damage assessment; Deep learning; Google street view; Remote sensing