| More than wide range of geological disasters in our country,and the landslides hidden,complexity,outburst,the uncertainty of space and time and dynamic change is strong.The frequency and amount of geological disasters are still at a higher level,the situation of geological disasters control is still severe.Landslide,as the second largest form of geological disaster after earthquake,causes a series of costly losses such as casualties,infrastructure damage,cultivated land damage,building collapse and so on.It is urgent to take relevant measures and technologies to effectively curb the frequent occurrence of landslide disasters,improve the ability of meteorological risk warning and prediction of landslide disasters,and effectively improve the ability of people’s life safety protection.In this thesis,fully promoting the development and deep integration of machine learning in the field of remote sensing application,so as to effectively improve the dynamic and rapid extraction technology level of landslide disaster,which is an important technical change in the traditional way of obtaining landslide spatial information.Based on XGBoost(Extreme Gradient Boosting),GBDT(Gradient Boosting Decision Tree),Ada Boost(Additive Boosting)and Light GBM(Light Gradient Boosting Machine)in integrated learning series algorithms,this thesis selected Fengjie County as the study area to study the rapid extraction of landslides from high-resolution remote sensing images.The specific research content is as follows:(1)A multi-scale segmentation parameter automatic optimization method based on landslide object was constructed.Based on the e Cognition software and with the help of the optimal segmentation scale estimation ESP2 tool,a multi-scale segmentation optimization method with landslide object as the core was created in this study,which realized the automatic optimization of landslide segmentation scale and ensured the integrity and definition of landslide object.(2)The Opt XGB landslide rapid extraction model was proposed.According to the spatial morphology and image characteristics of Fengjie landslide,8 landslide-prone towns in the northwest of Fengjie County were taken as test area.Feature optimization based on SHAP interpretation framework and automatic optimization method of Bayes superparameters based on Optuna framework were introduced into XGBoost basic algorithm.Then,Opt XGB,a rapid landslide extraction model,was constructed and proposed.Compared with the Opt GBDT,Opt Ada B and Opt LGB models optimized by the same method,the results verify the effectiveness of the proposed Opt XGB model for landslide extraction in mountainous cities and complex background remote sensing images.The accuracy rate of landslide extraction in the experimental area is larger,and the error rate and leakage rate are smaller.On this basis,Opt XGB model was successfully used to extract the spatial and temporal distribution information of landslides in Fengjie County from 2013 to 2020.(3)A rainfall-type landslide meteorological early warning model was established.The extraction of landslide temporal and spatial information based on high-resolution remote sensing images is to master and analyze landslide disasters that have occurred in the past.Further exploring the internal relationship between rainfall and historical landslides is helpful to timely warn the occurrence of future landslides.In this study,three critical rainfall threshold models of EI-D,EE-D and EE-EI were constructed on the basis of effective rainfall,and the accuracy of the three models was verified.The critical rainfall threshold model with the best accuracy was selected as the macro-meteorological early warning model of rainfall-type landslide in Fengjie County.On this basis,the classification form and corresponding warning measures of landslide meteorological classification in the study area were established. |
PDF Full Text Request