Rapid Damage Assessment Of Extreme Earthquake Disaster Areas Based On Remote Sensing Data

Remote sensing technologies can provide continuous data support for postearthquake emergency response efforts by observing the post-earthquake area on a large scale.In the age of remote sensing big data,the demand for remote sensing data information support for post-earthquake emergency work is also expanding in the direction of enhanced accuracy and refinement.In the National Earthquake Science and Technology Development Plan(2021-2035),it specifies that China should achieve international advanced status in earthquake emergency response science and technology by 2035.In order to achieve this goal,remote sensing technology is of immense importance.The main objective of this thesis is to investigate the rapid assessment of extreme disaster areas using remote sensing data.Based on the demands of earthquake emergencies and the rapid development of artificial intelligence technologies,the purpose of this study is to analyze the factors that limit the application of remote sensing data,and then to improve,develop and propose an intelligent analysis method and workflow for disaster identification in extreme earthquake disaster areas,and to enhance the postearthquake assessment of working patterns based on remote sensing data for extreme earthquake disaster areas.The thesis mainly completed the following work:(1)Review the development of post-earthquake remote sensing earthquake damage loss assessment,and remote sensing earthquake damage prediction methods;Theories,methods,and models for transfer learning and few-shot problems pertaining to information extraction.(2)To identify earthquake damage information in SAR images,we first proposed a collapsed object identification method based on the Light GBM method and massive SAR image features,which was verified using the Yushu earthquake in Qinghai.According to transfer learning theory,it is proposed to assign historical earthquake cases and new earthquakes to the source domain and the target domain in the transfer learning category,to define the source seismic event and the target seismic event,and then to apply joint distribution adaptations.Consequently,it was successfully transferred the seismic damage extraction model and knowledge from the source earthquake to the target earthquake,and sample and model reuse was achieved.Additionally,a method based on superpixel segmentation guidance and co-training is proposed for extracting earthquake damage information from single-phase polarimetric SAR data for post-earthquake conditions that lack samples.As a result,classification efficiency is improved.(3)For optical images,a change detection model of Siamese Network is built in a progressive way with domain adaptation capabilities.The model is shown to be capable of recognizing both single and group building collapses.A few shot semantic segmentation method is applied to the extraction of building earthquake damage from single-phase high-resolution optical images after the earthquake.And,an identification workflow for earthquake damage targets is carried out.(4)Improved the process of defining the area of extreme earthquake disasters and assessing loss caused by earthquakes using remote sensing techniques.The proposed approach is to identify the types and spatial distribution of disaster-bearing bodies through multi-source remote sensing data before the earthquake,and to combine multi-source information to delineate the scope of the extreme disaster area and assess the loss;after the post-earthquake images have been obtained,the loss assessment is revised based on the estimation of the earthquake damage.