| Undoubtedly,among all the natural disasters,the earthquake caused the most casualties,more than 50%of which are suffered by China.After the destructive earthquake,rapid,accurate and scientific determination of the location,scope and damage degree of earthquake-stricken areas is an important prerequisite for improving the efficiency of disaster relief,and is also the key to the success of emergency rescue decision-making.The development of deep learning,computer vision technology as well as the widespread availability of inexpensive unmanned aerial vehicles(UAVs)has provided a new approach to the rapid assessment of the severely earthquake-damaged area.On the basis of the integration of aftershock high-resolution unmanned aerial vehicle optical remote sensing images and building vector data acquired in the first National Geography Census in China,taking the full advantage of both the hierarchical feature extraction and transference abilities of the convolutional neural network(CNN),this paper focuses on the detection of the earthquake-induced building damage and tries to make a balance between the detection precision and the detection speed,aiming for a more rapid assessment of the severely earthquake-damaged area.From the aspects of precision and speed,this paper comes up with two novel models for building damage detection and deeply discuss the utility of the two models with the help of transfer learning in a coming earthquake.Eventually,an emergency-rescue-oriented method characterized by collapse rate calculation and multi-source data fusion for the rapid location of the severely earthquake-damaged area is proposed after three key questions,namely "Where are the houses?","Where are the collapsed houses?",and"How about the collapse rate?",are answered.The research in this paper aims to support the decision-making such as rescue team dispatch,relief material distribution and offer technical support for governments related to emergency rescue and disaster relief.The main researches are listed as below.(1)Based on the previous research achievements,considering the corresponding both national standards and industry standards comprehensively,the characteristics of the earthquake-induced building damage are analyzed and summarized.The structure of the damaged buildings is classified into two types,namely damaged buildings with wood structure and damaged buildings without wood structure.Besides,the seismic damage grades of the buildings are divided after the analysis of the typical damages.(2)A dataset for building damage detection is established on the basis of the highrevolution UAV images derived in ten days after Wenchuan Ms8.0 Earthquake,including two sub-datasets,namely a dataset with structure information and a dataset without structure information,consists of 5904 samples with 21047 damage instances.(3)Taking advantage of the Efficient Channel Attention Mechanism and feature fusion,a model named Earthquake-induced Building Damage Single Shot MultiBox Detector with Attention and Feature Fusion(EBD AFSSD for short)is proposed.The building damage is classified and located by means of the integration of both the shallow-level detailed features and the high-level deep semantic features and the feature enhancement made by the Attention Mechanism.The experiments prove that in comparison of the original Single Shot MultiBox Detector(SSD for short)model the detection ability of the model EBD AFSSD is largely improved with a mAP value up to 88.42%,4.55%better.More specifically,the mAP of the building damage with wood structure is 85.96%and that of the building without wood structure is 91.15%.Furthermore,in view of the lower requirement of the location precision of the building damage,a few of experiments are carried out in this paper to determine the best IOU value for the rapid assessment of the severely earthquake-damaged area.The results reveal that the model behaves the best when the value of IOU is set as 0.4.(4)In view of the strong timeliness of the emergency rescue,a building damage detection model with light weight named Earthquake-induced Building Damage SSDLite with Attention and Feature Fusion(EBD_AFSSDLite for short)is developed using the SSD framework with MobileNet V2 as its base net and also the attention mechanism.EBD AFSSDLite,merely 1/5 of SSD in model size,is 5 times faster and the mAP is up to 87.92%,1.46%better.It is worth mentioning that a 40000-iter training only cost about 6.5 hours but its mAP is up to 86.63%,which largely improve the timeliness of the model training.(5)In order to fully make use of the existing building damage samples and the pretrained models in a coming earthquake,this paper employs transfer learning and applies the method with an integration with building vector data acquired in the first National Geography Census in China in both the Wenchuan Ms8.0 earthquake and the Ludian Ms6.5 earthquake.It proves that an detection accuracy of above 90%can be achieved by sharing the pretrained weights of the middle,lower layers or all the layers obtained by the two models proposed in this paper,which fulfills the accuracy requirements of the rapid assessment of the severely earthquake-damaged area.In the case of Ludian Earthquake,49 more samples are made and the 132-epoch training only costs less than half an hour,which also fulfills the timeliness requirements.(6)This paper comes up with a method of rapid assessment of the severely earthquake-damaged area for emergency rescue on the basis of the calculation of collapse ratio of buildings in quake-hit regions.Specifically speaking,in combination with the building distribution obtained from the building vector data mentioned above,earthquake-induced building damage is located by directly detecting the original images with no mosaic.And then the collapse ratio is calculated.Besides the collapse ratio,the severely earthquake-damaged area is finally located by taking other evaluation basis such as earthquake influence field,seismic instruments intensity,ShakeMap,aftershock distribution focal mechanism solution and so on into account. |
PDF Full Text Request