Mbda-net: a building damage assessment model based on a multi-scale fusion network.

During or after a natural disaster, obtaining accurate disaster information and swiftly rescuing lives are of paramount importance. The objective of building damage assessment (BDA) is to use pre-disaster and post-disaster satellite images to predict the extent of building damage. In recent years, as the field of artificial intelligence advances, employing deep learning methods for BDA has gradually become mainstream. However, many existing methods use only pre-disaster and post-disaster images as inputs without considering their correlation. In this paper, we propose a model called building damage assessment model based on a multi-scale fusion network. In the first stage, we use a U-Net-based segmentation network to extract building location information from pre-disaster images, and the obtained network weights are then used for subsequent assessment. In the second stage, we adopt a dual-branch U-Net as the backbone network, which inputs both pre-disaster and post-disaster images. We design the feature multi-scale fusion module to extract and integrate multi-scale features, and design the cross squeeze-and-excitation module to analyze the correlation between pre-disaster and post-disaster images. Our method is validated on the xBD and LEVIR-CD datasets, and achieving F 1 b = 0.862 for segmentation metrics and F 1 d = 0.759 for classification metrics. [ABSTRACT FROM AUTHOR]