Research On Debris Flow Susceptibility Evaluation Based On Stacking Ensemble Learnin

Debris flow is a common geological disaster in mountainous areas,which is the result of a combination of factors and often has the characteristics of suddenness,rapid flow and strong destructive power,and is second only to earthquakes in terms of danger and impact.Dongchuan district is located at the northern edge of Yunnan-Guizhou Plateau,and its natural environmental conditions,meteorology,hydrology and geological environment are complex and diverse.Its special topographic and climatic conditions and human activities cause frequent geological disasters in the whole study area,among which debris flow disaster is the most serious,and it is known as a natural museum of debris flow in the world.The assessment of debris flow susceptibility is an effective way to prevent and control debris flow disasters,so that decision makers can better understand the spatial probability of debris flow disasters in Dongchuan District,and then monitor and predict debris flow disasters more accurately and provide timely warnings to minimize the damage caused by mudslide disasters.Therefore,only based on various types of spatial data in the region,the comprehensive use of geographic information science,computer science and other innovative methods,in-depth study of the spatial distribution of regional debris flow susceptibility,in order to provide theoretical and methodological support for the later prevention and mitigation of disasters.With the development of computer technology,many machine learning algorithms are widely used in the field of debris flow disasters,but the susceptibility of debris flow is a complex nonlinear problem,and the limitations of a single learner cause the constructed susceptibility model to have limited generalization ability.The accuracy and generalization ability of the model can be significantly improved by coupling multiple classification learners,and this method has been widely used in many fields,but there are few related studies in the field of debris flow hazard susceptibility assessment.Therefore,based on the interpretation of debris flow points and based on multi-source data such as landform,geology and precipitation,this paper selects 13 debris flow pregnancy factors,and uses the frequency ratio model to analyze the relationship between each pregnancy factor and the distribution of debris flow.At the same time,it uses the Spearman correlation coefficient and information gain ratio to analyze the correlation between each pregnancy factor and its contribution to the development of debris flow.Finally,for the first time,we combined GIS and Stacking integrated learning framework algorithm to study the debris flow susceptibility.The main results obtained are as follows:(1)13 debris flow pregnancy factors,including elevation,slope,slope direction,plane curvature,profile curvature,terrain moisture index(TWI),May-October precipitation,distance from river network,distance from road,distance from fault,normalized vegetation index(NDVI),stratigraphic lithology and land use type,were selected from five aspects: topography and geomorphology,hydrometeorology,geological conditions,vegetation cover and human activities.The frequency ratio model was used to analyze the relationship between the disaster pregnancy factors and the distribution of debris flow.The results show that land use type,stratigraphic lithology,NDVI(0.0869 ～ 0.2335],elevation less than 1519 m and distance from fault(1056 m,2112 m] are most closely related to the distribution of debris flow,and other meteorology and hydrology and human activities were also more related to the distribution of debris flow.(2)Based on the analysis of the relationship between each pregnancy disaster factor and the distribution of debris flow,the correlation and contribution rate of each pregnancy disaster factor are analyzed by using the Spearman correlation coefficient and information gain ratio.The results show that there was no obvious correlation between the disaster pregnancy factors,while the four factors of NDVI,profile curvature,May-October precipitation and plane curvature had the largest contribution to the development of debris flow,among which the distance from the road network and slope had the lowest contribution,and all the 13 selected factors had different degrees of contribution to the development of debris flow.(3)Support vector machine(SVM),BP neural network and classification regression tree(CART)as base learners and logistic regression(LR)as meta-learner were used to construct debris flow susceptibility assessment model(SBC)in Dongchuan district based on Stacking integrated learning framework for the first time,and the debris flow susceptibility indices were calculated under different models.The susceptibility index was classified into five susceptibility classes,namely,very high susceptibility,high susceptibility,medium susceptibility,low susceptibility and very low susceptibility,using two methods: natural breakpoint method and K-mean clustering,and finally the susceptibility class map of the study area was drawn.The results show that the low and very low susceptibility zones are most widely distributed in Dongchuan District,mainly in Hongtudi,the southern part of Shekuai,Inmin and Awang,and the eastern part of Tongdu Street;the very high and high susceptibility zones are mainly distributed in the two banks of Xiaojiang River and the south bank of Jinsha River,where the geological environment is fragile and the risk is high.(4)By statistically calculating the assessment index of each model and the difference of mudslide density between very high and very low susceptibility levels of each model.It was found that the AUC,ACC and F1 scores of the SBC model constructed based on the Stacking integrated learning framework were 96.65%,91.07% and 91.30%,respectively,with an average increase of 3.03%,4.19% and3.77%,respectively,compared with the other three single models;the SBC model constructed based on the Stacking integrated learning framework had The difference of debris flow density of the SBC model constructed based on Stacking integrated learning framework are all higher than the other three single models.This indicates that the SBC model constructed with the Stacking integrated learning framework is more scientific and more accurate than the other three single models in assessment susceptibility.