Research On Extraction Of Spatio-temporal Semantic Information And Topic Evolution Of Meteorological Disasters

In recent years,the frequent occurrence of global natural disasters has led to serious human economic loss and environmental damage,and the research on meteorological disaster events has attracted more and more attention.Social media platform contains a lot of information.This bottom-up information can be used as an alternative information source to help research disaster events when the traditional observation data is scarce.This study focuses on the extraction and analysis of spatio-temporal semantics of meteorological disasters,using the crawling micro-blog data of the main urban area of Hefei with geographical labels from June 19 to July 15,2016 and the microblog data of typhoon “Mangosteen” without geographical labels from 00:00 on September 16 to 09:10 on September 27,2018.Combined with DBSCAN(Density-Based Spatial Clustering of Applications with Noise)spatial clustering algorithm,LDA(Latent Dirichlet Allocation)topic model,LSTM(Long Short-Term Memory Neural Network)classification model and Markov transition probability matrix,information extraction and topic evolution analysis are carried out in order to assist decision makers to understand the disaster situation in time and provide emergency response reference in reducing the damage caused by natural disasters.The main conclusions of this study are as follows:(1)From the perspective of time,the rainstorm is closely related to the change of the number of micro-blogs.The occurrence of rainstorm will promote the lag growth of the number of micro-blogs published,but the long-lasting growth of the number of heavy rainfall micro-blogs is not obvious.Spatially,the rainstorm micro-blog is mainly distributed in large-scale comprehensive areas with developed economy,and at the same time,residential communities and other traffic intensive areas.(2)The distribution and evolution of the topic of rainstorm micro-blog in different time stages and different geographical spaces have obvious characteristics.Through the DBSCAN clustering of micro-blog data,15,5 and 8 subcategories were obtained in three periods,and LDA model was introduced to mine the topics of each spatial sub-category.In the early stage of rainstorm,it is suggested to make preparations for traveling in the rain,pay attention to the traffic inconvenience caused by surface water,and worry about the possible flood disaster in some areas in the future.In the middle stage of the rainstorm,the sustained heavy rainfall caused huge economic losses,leading to serious floods and other derivative disasters,so the flood fighting and emergency operations are imminent.In the late stage of rainstorm,the weather is getting better gradually,and the reconstruction is particularly important.People focus on the post disaster reconstruction work organized by the government,such as the excavation of buried roads,the flood relief after rain and the current situation of some areas.(3)Using LSTM to train the subject classification model of typhoon microblog,the accuracy of the model is more than 85%.The training effect of the model is good,and still has some space to optimize.Typhoon micro-blog is divided into six topics: "disaster prevention and rescue","transportation and public facilities","affected people","others","weather and early warning" and "defense and emergency response".The number of micro-blogs on "transportation and public facilities" is more than a quarter of the total.There were 6409,7638 and 4631 typhoon micro-blogs in the early,middle and late stages respectively,with the largest number of micro-blogs in the middle stage.Before typhoon landing,the main content of micro-blog was "weather and early warning".After landing,the proportion of "weather and early warning" continued to decline.In contrast,the percentage of "traffic and public facilities" micro-blogs continued to rise after the typhoon landed,exceeding one third of the total number of micro-blogs in the late stage.(4)Through word segmentation statistics and word cloud map to aggregate the information,we can extract the hot events of six kinds of topic micro-blogs and mine the topic content.The visualization of the key words in the word cloud helps to display the information related to the typhoon disaster,including the casualties,the location and disaster situation of the affected area,the derivative disasters caused by the typhoon,the emergency response measures implemented by the relevant national departments,the public’s views and discussions on the typhoon,and the impact of the typhoon on people’s transportation and daily life.(5)Using Markov probability transition matrix to calculate the time conversion of typhoon specific tpopic categories,compared with the transition stages from the early to the middle and from the middle to the late,"transportation and public facilities","other" are the main transformation directions of micro-blog topic.The damage of typhoon and its derivative disasters to traffic environment and public facilities makes "traffic and public facilities" occupy the leading position all the time.In the middle stage,the masses had a strong response to the typhoon landing and made a large number of relevant comments for discussion,which led to an increase in the probability of "other" topic transition.The great negative impact of late typhoon landing on the public has enhanced the topic concern of the "affected people ".