Severe earthquakes always lead to catastrophic building damage. Post-earthquake building damage level classification (BDLC) is an important task to rescue persons and make rapid earthquake responses for the reduction of severe injuries and casualties. To reduce data processing time for post-earthquake disaster response, pre-earthquake building data are always prepared, because pre-existing information about building locations and characteristics can reduce the time of post-earthquake localising buildings. Therefore, both pre-earthquake building information preparation and post-earthquake building damage information collection facilitate swift BDLC. Compared with conventional labour-intensive, time-consuming, and possibly dangerous in-situ observations, remote sensing technology provides a rapid and efficient approach to these pre- and post-event data collections because of its capability to acquire large-scale data remotely and rapidly. There are several remote sensing data types with their own advantages. For instance, two-dimensional (2D) optical satellite images provide large-scale information of the earth. Three-dimensional (3D) Light Detection and Ranging (Lidar) point clouds, as another type of remote sensing data, provide additional information on elevations of ground and non-ground points, including the heights of buildings.

Among the methods for processing 2D and 3D remote sensing data, deep learning semantic segmentation (DLSS) technology has a high potential in applications for BDLC on remote sensing data. However, the potential of these methods for BDLC has not been thoroughly studied in previous research. Indeed, there are four gaps in the literature in this domain. Firstly, few DLSS methods have been applied to 2D satellite images or 3D point clouds for building damage classifications specifically related to earthquakes. Secondly, several well-known DLSS algorithms were proposed and tested only on small or indoor case studies in 2D and 3D applications. The large-scale outdoor scenarios have yet to be fully discussed or tested. Thirdly, most current post-earthquake BDLC studies lack detailed multi-level classification methods in the remote sensing field. Fourthly, for the training of the DLSS methods, there is a lack of labelled datasets for multi-level BDLC at large study extents in pre- and post-earthquake events.

This study solved these problems by applying these methods to both 2D and 3D remote sensing data on large-scale outdoor areas and by proposing novel DLSS approaches to classify building damage into four levels. To overcome the lack of training data, this study prepared and developed labelled datasets for the training of the proposed DLSS methods.

Ablation studies have been designed to test the performance of these proposed DLSS methods. The results in this study show the good performance of these methods at large-scale building footprint extraction and four-level BDLC with either satellite or Lidar data. Indeed, these novel methods have increased the accuracy of the chosen backbones in large-scale outdoor study areas. The channel attention mechanism helps to improve the accuracy of building information extraction in both 2D and 3D methods with higher Intersection over Union (IoU) values compared to the chosen backbones. Overall, this study overcomes the issues of the current methods of BDLC and will benefit society by providing a safe and speedy post-earthquake BDLC method that requires minimal fieldwork to support rescue teams for quick response. It will also help disaster management systems to store information efficiently for post-earthquake recovery planning.