Multi-source Domain Adaptation And Its Application In Post Disaster Building Damage Detection

In recent years,global disasters have occurred frequently,causing heavy losses to the safety of human life and property.After a disaster occurs,it is extremely important to obtain accurate information on building damage in a timely manner for formulating rescue strategies and carrying out rescue work.Given that it is difficult to quickly obtain labeled samples in the area after a disaster,and it is impossible to train a deep learning model with good performance,transfer learning has gradually attracted attention.However,the data distributions in two different scenarios are different,and simply applying the model trained in one scenario to another scenario to perform classification tasks will seriously affect the transfer performance.Therefore,how to transfer the knowledge learned in the historical scene to the current scene and make use of it has become a hot research topic today.In response to the difference between data distributions,researchers have proposed the concept of domain adaptation to solve the problem of migration when the tasks are the same and the data distributions are similar.However,the traditional domain adaptation method assumes that the training data is sampled in a single domain,ignoring the actual situation of collecting labeled samples from multiple sources,so there is an extended method called multi-source domain adaptation(MDA),fully excavate the effective information in the historical scene,and improve the transfer efficiency of knowledge.However,due to the variety of architectural styles in different regions,there are not only differences between historical scenes and current scenes,but also differences between historical scenes,which will have a negative effect on knowledge transfer,so MDA is needed to accomplish the task of eliminating the two differences.According to the timing of fusion between different source domains,MDA can be divided into two ideas: "first fusion" and "post fusion".However,the first fusion MDA needs to improve the effect of generating images,and the difficulty in determining the weights in the fusion method leads to relatively complicated operations.In response to these two problems,this paper proposes a multi-source adversarial domain adaptation(MADA)and multi-source variational domain adaptation(MVDA)algorithms are used to evaluate the damage of buildings after disasters.MADA is divided into two steps:creating the adaptation source domain and adapting the feature-level domain adaptation between the source domain and the target domain.MVDA consists of two stages: the feature-level domain adaptation between each source domain and the target domain and the classifier alignment.The two methods make full use of the effective information in multiple historical scenarios,solve the problem of mutual interference between historical scenarios,and improve the migration efficiency from historical disaster scenarios to current disaster scenarios.Two experiments are conducted with the Hurricane Sandy,Irma,and Maria datasets as multi-source and target domains to verify the effectiveness of MADA and MVDA.The results show that the proposed method outperforms other compared methods in classification performance.