Discrimination of Earthquake-Induced Building Destruction from Space Using a Pretrained CNN Model

Abstract

The building is an indispensable part of human life which provides a place for people to live, study, work, and engage in various cultural and social activities. People are exposed to earthquakes, and damaged buildings caused by earthquakes are one of the main threats. It is essential to retrieve the detailed information of affected buildings after earthquakes. Very high-resolution satellite imagery plays a key role in retrieving building damage information since it captures imagery quickly and effectively after the disaster. In this paper, the pretrained Visual Geometry Group (VGG)Net model was applied for identifying collapsed buildings induced by the 2010 Haiti earthquake using pre- and post-event remotely sensed space imagery, and the fine-tuned pretrained VGGNet model was compared with the VGGNet model trained from scratch. The effects of dataset augmentation and freezing different intermediate layers were also explored. The experimental results demonstrated that the fine-tuned VGGNet model outperformed the VGGNet model trained from scratch with increasing overall accuracy (OA) from 83.38% to 85.19% and Kappa from 60.69% to 67.14%. By taking advantage of dataset augmentation, OA and Kappa went up to 88.83% and 75.33% respectively, and the collapsed buildings were better recognized with a larger producer accuracy of 86.31%. The present study showed the potential of using the pretrained Convolutional Neural Network (CNN) model to identify collapsed buildings caused by earthquakes using very high-resolution satellite imagery.

Multi-Hazard and Spatial Transferability of a CNN for Automated Building Damage Assessment

“…To know if the model indeed shows this behavior, the performance of all classes should be analyzed separately. Only two studies reported this, and in both cases the model performed better on the majority class (not destroyed) than on the minority (destroyed) class [32,45], suggesting that a larger class imbalance can lead to better accuracy while not indicating a better predictive value.…”

“…Combined they covered three types: earthquake, tsunami, and hurricane, where earthquake is the most common type. Among all studies shown in Table 1, those focusing on earthquakes [13,32,37,44,45] reported the lowest performance. However, since all models have only been tested on one hazard type, it should be further researched if and how the disaster type influences the performance.…”

“…At the same time, the quality of the different models cannot be directly compared based on the reported accuracy, since there are more differences between the studies than solely the model; these include class balance, disaster type, and data sources. We also note that over the years models have become more complex [13,32,[42][43][44][45], but this has not led to an evident increase in accuracy, as shown in Table 1.…”

“…As a source for the damage labels, most studies used the analysis done by UNITAR/UNOSAT [32,44,45]. These data have been created through the annotation of satellite imagery by professional analysts and are openly available through the Humanitarian Data Exchange (HDX) (https://data.humdata.org).…”

“…Most studies worked with an imbalanced dataset with respect to the difference between the number of destroyed and not destroyed buildings [32,42,44,45], as shown in Table 1. The two studies that had a balanced dataset [13,43] prepared this on purpose by selecting solely a part of the original data, and the same was done, to a certain extent, by the other studies to diminish the imbalance.…”