Research On Post-earthquake Building Collapse Identification Based On Swin Transformer Network Architecture

In recent years,earthquake disasters have become one of the most serious natural disasters in the world.China has a vast territory and extremely complex geological structure.Many regions are prone to frequent earthquakes,which seriously threaten people’s lives,property safety,and social and economic development.After an earthquake occurs,timely assessment of the seismic damage to urban buildings is crucial for government emergency decision-making and post-disaster rescue.Traditional manual inspections have limitations such as low timeliness and poor accuracy,making it a long-standing challenge in the field of engineering structure disaster prevention and reduction to obtain accurate building damage information in a timely,accurate,and comprehensive manner.Satellite optical remote sensing images have become one of the main information acquisition methods in the early stage of earthquake relief work due to their multi-angle,all-weather imaging characteristics.However,traditional image processing methods for interpreting remote sensing images are not only inefficient and inaccurate,but also have weak generalization ability.The Transformer deep learning model with anti-interference ability and robustness has the ability of global analysis and multi-level feature extraction of images,providing a new idea for the extraction of seismic damage information of building groups in remote sensing images.In this research,we propose a method for identifying urban building group seismic damage based on an enhanced Swin Transformer network architecture,focusing on common challenges such as dense distribution of buildings,strong weather interference,large geometric shape differences,different resolutions,and diverse data sources in post-earthquake remote sensing images.We conducted a systematic pixel-level semantic segmentation task for different source data,and verified the effectiveness of the model method through actual earthquake data.The main research contents are as follows:(1)To address common challenges in remote sensing images such as complex backgrounds and strong weather interference,a data and feature information enhancement processing algorithm of brightness transformation and fogging is used to simulate the possible situations of darkness,overexposure,and cloud and fog obstruction in real remote sensing images,and to systematically train the network model under multiple working condition scenarios.(2)An enhanced Swin Transformer network architecture model with an encoder-decoder structure is proposed,which achieves multi-level feature fusion by adding a convolutional feature fusion module,and inserts a channel attention and spatial attention module(CBAM)into the model to enable the network to focus on key feature information,while using UPer Net as the decoder output recognition structure.The optimized model achieves an m Io U accuracy value of 88.53% on the validation dataset of remote sensing images in Yushu and Beichuan,a 1.3%improvement compared to the original network.The testing accuracy values on the test dataset with strong weather interference in the two regions are 72.14% and89.91%,respectively.Through ablation experiments,the mechanism of feature fusion and CBAM modules and the superiority of the enhanced Swin Transformer model compared to multiple CNN-based segmentation models are verified.To address the need for improved generalization and robustness of the model training due to the limited availability of satellite high-altitude remote sensing data,a dataset of fused high-altitude satellite and low-altitude drone multisource and multi-resolution remote sensing data is established,and an adaptive feature processing module is introduced.By optimizing the hyperparameters of the enhanced Swin Transformer network model,an m Io U value of 90.11% is achieved on the validation dataset.In addition,the optimized model achieves m Io U values of 79.97% and90.95% on the satellite remote sensing image test datasets with weather interference in Yushu and Beichuan,and 85.42% on the drone remote sensing image test dataset with weather interference,effectively achieved the recognition and classification of building collapse based on multi resolution remote sensing images after the earthquake.