This section discusses the methods that recently were used for building damage assessment using remote sensing satellite images. Due to the high impact of natural disasters on human lives, it has always been a hot topic to study various techniques to cope with the damages that occur to cities. Many researchers from two decades ago have started to use satellite images for the purpose of building damage assessment [6,20,42]. In the meantime, researchers have developed more sophisticated methods based on machine learning techniques to automate the assessment process [9,25,43]. As an example, Putri et al. [25] tried to generate building damage assessment maps using textural and spectral features extracted from pre- and post-disaster Sentinel-1 and Sentinel-2 images. They fed the features into a Random Forest classifier for classification. Alatas et al. [44] compared different morphological profiles with Haralick’s texture features to detect damaged buildings using a k-Nearest Neighbor classifier. Likewise, Janalipour et al. [45] utilized an ANFIS-based decision-making system powered by geometrical features and bi-temporal satellite images combined with building vector maps to detect damaged buildings in the Bam earthquake.

Recently, deep learning techniques have been widely studied to provide more general solutions for the task of building damage assessment [19,36]. The ability of deep learning models to provide high-dimensional representations of images has enabled them to solve more sophisticated challenges. For instance, Zheng et al. utilized a deep object-based semantic change detection method to evaluate damages to buildings that occurred due to natural and man-made disasters [19]. In [46], Wu et al. proposed a variety of attention UNets to localize and classify building damages. By detecting changes in super-pixels, Qing et al. [14] proposed a CNN-based network for damage assessment using UAV images. Moreover, in [12], Deng et al. proposed a two-stage UNet supplemented with the Squeeze and Excitation (SE) module to improve the damage assessment results.

According to the usage of only post-disaster images [11,32,47] or the pair of pre- and post-disaster images [12,34,36], various methodologies have been proposed for building damage assessment. In [29], Duarte et al. proposed various fusion strategies for building damage assessment from post-disaster satellite, aerial, or UAV images using CNNs. In [15], Tilon et al. proposed an unsupervised method for building damage assessment. However, without any guidance from pre-disaster images or possible external sources of building footprint maps such as OpenStreetMap [15], post-disaster images cannot provide precise building boundaries [14]. In this regard, in 2013, Dong et al. [9] showed that most of the studies tend to use pairs of pre- and post-disaster images to be able to locate the damaged building and evaluate the damage by detecting meaningful changes. Accordingly, Khodaverdizahraee et al. [48] extracted building properties such as shape, geometry, shadow, texture, and color from pre- and post-disaster imagery and fed them into machine learning algorithms for improved building damage assessment. Likewise, Xu et al. [49] compared different pre- and post-disaster feature fusion scenarios in CNN models for the Haiti earthquake.

From the architecture’s point of view in deep learning methodologies, various architectures from the computer vision community have been proposed for building damage assessment. Inspired by atrous spatial pyramid pooling networks, with dilated convolutions, Gupta et al. [50] proposed a network (RescueNet) with multiple segmentation heads for simultaneous building detection and damage classification. Valentijn et al. [51] studied the effective parameters of damage assessment predictions by using simple, fully connected CNNs. Compared with other architectures, UNets, with encoder-decoder paths that were improved by adding skip-connections to aggregate low-level and high-level features, have been widely used for the task of building damage assessment [12,36,46].

In order to improve the performance of baseline semantic segmentation architectures, researchers have used various strategies and extensions in their networks. For instance, Bai et al. [52] used a lightweight approach based on knowledge distillation to reduce the dependence on computing resources and increase the speed of damage assessment in emergency situations. In contrast, in order to achieve more accurate results with fewer labels, Qiao et al. [53] implemented a weakly supervised technique that could improve the quality of activation maps and boost model performance. Attention techniques [34,53], which try to emphasize important regions of the images, have been widely used in computer vision tasks. Mei et al. [34] used a difference-aware attention strategy to predict more accurate localization and classification results after disasters. For a superior combination of features in spatial and channel directions, Shen et al. [36] proposed BDANet with a cross-directional attention mechanism and obtained enhanced damage assessment maps. Attention-based mechanisms have shown promising results in both computer vision and building damage assessment.

From the development point of view, appropriate and carefully annotated datasets play a major role in deep learning models. Deep learning models are data-hungry, and the issue is exacerbated by the complexity of the problem [54]. Building damage assessment is a complicated problem and, thus, requires large amounts of ground truth data for model development. Preparing annotated remote sensing images for such tasks is time-consuming and requires huge manual work. Although the number of benchmark datasets is growing in the remote sensing community, high-quality datasets that are suitable for building damage classification are still rare [55]. The xBD dataset [56] has played a game-changing role in building damage assessment during the past five years. However, it still has some challenges, such as highly biased classes towards no-damage buildings. Furthermore, complexities of minor-damage and major-damage classes, besides similarities between minor-damage and no-damage classes, lead to moderate misclassification [57]. Preprocessing techniques such as data augmentation have been widely utilized to increase the size of data and robustness of the model [58]. They can also act like regularization techniques that feed models with more challenging scenarios [58,59]. Advanced augmentation methods, such as the Cutout, randomly mask out square regions of input images to improve the feature representation ability of the model [41].