Situation Awareness Of Urban Rainstorm Disaster Based On Topic Evolution And Multimodal Feature Fusion Of Social Media

In the context of global climate change,there are more and more extreme precipitation events in China.At the same time,the flood defences in cities have not been improved accordingly.Affected by rainstorms every year,a large number of economic losses and casualties occur,which directly affects China's economic development and people's property security.In rainstorm disaster management,the government and volunteer rescue organizations need to collect and understand as much relevant information as possible,that is,situational awareness,which is essential for the implementation of emergency response activities such as disaster reduction and rescue.Because of the complexity of disasters,it usually takes a lot of human and financial resources to gain large-scale situation awareness at the city level in a short time.With the development of the mobile Internet,various social networking sites such as Twitter,Facebook,Sina Weibo,etc.have developed rapidly.The rainstorm disaster researchers and managers have tried to use social media data for situational awareness.However,current researches mainly has the following problems:(1)Social media-based situational awareness research is mostly based on social media text mining and classification.However,the ambiguous semantics of social media short text,making it difficult for text classification and topic mining.The current topic mining methods are hard to support subsequent spatiotemporal analysis.(2)In the researches of situational awareness that combines social media semantics and time information,the currently regularly used topic analysis methods can not consider the internal connections between topics.And so that,such methods are hard to establish the evolutionary relationship between topics,challenging to quantitatively analyze the relationship between the evolution process of social media topics and the stages of urban rainstorm disaster.(3)In the studies of situational awareness combining social media semantics and spatial information,the multi-modal characteristics of social media are ignored.Due to the uncertainty of social media text data,such text mining-based methods are challenging to support fine-grained spatial analysis,leading to difficulty for accurately identifying and assessing urban waterlogging.To solve the above three problems,this article mainly includes the following three aspects of work:(1)This paper proposed a new topic discovery method based on co-words(TDMBOC).And based on this method,the crisis information related to rainstorm disaster in social media data is mined.Specifically,we first proposed a method to extract topic words related to rainstorm disaster in social media data by combining term frequency–inverse document frequency(TF-IDF)and head/tail breaks algorithm.Based on the extracted topic words,two schemes for constructing co-word network in social media were proposed,which include the co-word network with documents as the node(DNCW-NET)and the co-word network with topic words as the node(TWN-CW-NET).Based on the two kinds of co-word network,we suggested using co-word network communities to represent topics in social media texts.We explored the applicability of these two co-word network-based methods for social media short text topic mining,and the conclusions are as follows: in terms of topic discovery,the two co-word network-based methods have higher accuracy and recall rate than LDA model;the topic discovery results based on DN-CW-NET is slightly better than that based on TWN-CW-NET;The social media topic mining results based on DN-CW-NET can be used to support the subsequent spatiotemporal analysis and mining.(2)In this paper,a community evolutional meta-network(CEM-Network)model is proposed.Based on this model,the topic evolution process in social media data and its relationship with the disaster management stages are analyzed.Specifically,we first proposed to use the community evolution of TWN-CW-NET to represent the topic evolution in social media.And then,we identified several types of community evolution events for analyzing topic evolution,designed a simplified group evolution discovery(GED)algorithm to detect these topic evolution events,explored the setting method of key parameters in the GED based on head/tail breaks algorithm.Based on the detected topic evolution events,we introduced the concepts of community evolutional hyper network,hypercommunities,and lifetime of hypercommunities to quantify the topic evolution in social media,and explored the correlation between the topic evolution in social media and the stages of disasters.This research has concluded that the community evolutional hyper network model based on TWN-CW-NET can effectively quantify the evolution of related crisis topics in social media data when a rainstorm occurs,revealing the stages and processes of the rainstorm disaster.(3)A method for urban waterlogging detection and assessment based on multimodal transfer learning is proposed.Specifically,We first developed a visual feature transfer module and obtained the visual features of the social media image modal by transferring a sizeable external image dataset.Then,we created a text sentiment transfer module and got the sentimental features of the social media text modal by transferring an external comment tendency dataset.Finally,A multi-modal feature fusion model was proposed for waterlogging assessment in urban rainstorm disasters.The conclusions of this research include: when using social media for situational awareness,especially when using social media for fine-grained spatial analysis,the social media's multi-modal information can effectively overcome the uncertainty of social media text modal data;the method proposed in this research can effectively fusion the multi-modal information in social media data,and accurately assess waterlogging in the rainstorm disasters.