Research On Identification Of High-risk Areas Of Flash Flood Disasters Based On Big Data And Machine Learnin

Mountain torrent disaster is a common destructive natural disaster in hilly areas.It has the characteristics of suddenness and difficulty in prediction.It seriously threatens the safety of people’s lives and property and is one of the important constraints on the high-quality economic development in hilly areas.Scientific risk identification can provide important technical support for flood management and disaster reduction.This study is based on the national natural science foundation program named “Identification of flash flood risk driving factors and simulation of amplification effect(42271095)”.Fujian Province,which is prone to mountain torrent disasters,is selected as the research area,and small watersheds are used as units.Based on the self-built geographic database,the risk identification of mountain torrent disasters is studied to provide an important reference for mountain torrent disaster prevention.The main research contents and results are as follows:(1)Create a database.Based on the results of the investigation and evaluation of mountain torrent disasters,collecting the information of mountain torrent disasters generated by Internet users,compiling according to the data structure,and the data processing work such as river network registration and contour layer drawing of residents along the river was carried out,the geographic database of mountain torrent disaster risk analysis in Fujian Province was created,which provided data support for this research and subsequent related risk research.(2)Spatio-temporal characteristics analysis and driving force analysis of mountain torrent disasters.The spatial and temporal pattern characteristics of mountain torrent disasters in Fujian Province from 1949 to 2021 were analyzed by M-K test,kernel density analysis,standard deviation ellipse and mean center space analysis.The results showed that: 1)the overall trend of mountain torrent disasters was on the rise after 1988;2)The highly concentrated areas of mountain torrent disasters are mainly located in Quanzhou and Ningde,and the focus of mountain torrent disasters with a research period of about 20 years from 1949 to 2021 has been located in Quanzhou.On the whole,the spatial change is from aggregation to diffusion,and the degree of dispersion is increasing and the direction is weakened.In addition,the relationship between mountain torrent disasters and influencing factors was explored,including the analysis of the characteristics of mountain torrent disasters and rainfall in the upstream basin and the response relationship between flood discharge capacity and mountain torrent disasters in small watersheds.(3)Construction of mountain torrent disaster risk assessment index system.The index selection focusing on the characteristics of runoff yield and confluence in small watersheds,combining the results of the relationship between mountain torrent disasters and influencing factors,according to the regional disaster theory,10 primary indicators are selected from the three aspects of disaster-causing factors,disaster-pregnant environment and disaster-bearing bodies.Then through the multicollinearity analysis and importance measurement of the indicators,the design rainstorm,24 H rainstorm variability,slope,area,shape coefficient,flood discharge capacity,river slope and residential area within 300 m of the upstream watershed3H10a(10-year return period 3-hour rainfall duration)are finally determined as evaluation indicators,and the mountain torrent disaster risk evaluation index system is constructed according to the hierarchical structure.(4)Construction of mountain torrent disaster risk identification model.The risk identification model is constructed based on random forest and support vector machine algorithm respectively,and the model is trained and verified by using historical mountain torrent disaster data.The model performance is measured by five indexes: accuracy,precision,recall,F1 score and area under ROC curve(AUC value).The results show that the performance of the random forest model is better than that of the support vector machine model,so the random forest model is selected to identify the mountain torrent disaster risk in Fujian Province.(5)Analysis of mountain torrent disaster risk identification results.Based on the random forest model,the risk distribution diagram of mountain torrent disaster is obtained.The accuracy of risk level is verified by superimposing historical mountain torrent disaster points.The results show that 89.38% of historical mountain torrent disasters fall into the higher-risk and high-risk areas of mountain torrent disasters,indicating that the classification of mountain torrent disaster risk identification results based on the index system considering the characteristics of catchment and random forest model is reliable.The areas with higher risk of mountain torrent disaster and above are roughly inclined “U”-shaped distribution in Fujian Province,and the risk level in the eastern region is high.Finally,this thesis also puts forward corresponding mountain torrent disaster prevention and control suggestions for small watersheds with different risk levels of mountain torrent disaster in the identification results.