Post-disaster building damage assessment based on improved U-Net

Abstract

When a severe natural disaster occurs, the extraction of post-disaster building damage information is one of the methods to quickly obtain disaster information. The increasingly mature high-resolution remote sensing technology provides a solid foundation for obtaining information about building damage. To address the issues with inaccurate building positioning in existing building damage assessment methods, as well as poor classification due to similar minor and major damage characteristics in building damage classification. Based on U-Net, we designed a two-stage building damage assessment network. The first stage is an independent U-Net focused on building segmentation, followed by a Siamese U-Net focused on building damage classification. The Extra Skip Connection and Asymmetric Convolution Block were used for enhancing the network's ability to segment buildings on different scales; Shuffle Attention directed the network's attention to the correlation of buildings before and after the disaster. The xBD dataset was used for training and testing in the study, and the overall performance was evaluated using a balanced F-score (F1). The improved network had an F1 of 0.8741 for localization and F1 of 0.7536 for classification. When compared to other methods, it achieved better overall performance for building damage assessment and was able to generalize to multiple disasters.

Bitemporal Attention Transformer for Building Change Detection and Building Damage Assessment

“…Siam-U-Net-Attn introduced a self-attention module to incorporate long-range information from the entire image [14]. Deng and Wang [17] used shuffle attention to correlate buildings before and after the disaster. LGPNet incorporated two general attention mechanisms, the position attention module and the channel attention module [34].…”

“…Deng and Wang [17] develops a two-stage BDA network based on the U-Net architecture. The initial stage employs an independent U-Net for precise building segmentation, succeeded by a Siamese U-Net dedicated to building damage classification.…”

“…However, existing BDA methods [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19] overlook the intrinsic order among ordinal labels and simplistically treat the BDA task as a multiclass semantic segmentation problem. These methods employ traditional classification loss functions, such as crossentropy loss, dice loss, and focal loss, to train their change detection networks.…”

DISCount: Counting in Large Image Collections with Detector-Based Importance Sampling

“…The second application is building damage assessment for disaster response ( § 4.2). Aid organizations assess building damage using satellite images (Deng and Wang 2022) to plan humanitarian response after a natural disaster. However, pre-trained detection models are often not accurate enough when applied to a new dis- aster due to domain shift, and there is very limited time and expertise for model development, but volunteers are available to view satellite imagery and help assess damage.…”

Deep Ensemble Learning for Rapid Large-Scale Post-Earthquake Damage Assessment---Application to 2023 Kahramanmaraş Earthquake Sequence

“…Most previous studies in this area fall into the categories of image segmentation [11,12] or object detection [13]. The insights offered include but are not limited to, the effectiveness of "attention mechanisms" and visual transformers built on those mechanisms [14,15,16,17,18], novel convolutional blocks [19] and even graph neural networks that account for dependencies between structure types and their damage conditions [20]. Other studies use additional data-for example, public building footprint inventories such as OpenStreetMap 2 for building localization [21], or as additional input data channels [22,23]-, or use both pre-and post-event images for damage identification through change detection strategies [24].…”