| The Loess Plateau has a variety of erosion processes due to its unique natural environment.As a micro-topography widely distributed in the Loess Plateau,the loess sinkhole is also a unique way of soil erosion and geological disaster type,and While affecting the development process of the Loess Plateau landform,it is very easy to cause serious soil erosion and damage to engineering facilities in the Loess Plateau.It will induce other types of geological disasters such as debris flow,landslide,and collapse,indirectly causing casualties and social economic losses,threatening the safety of people’s lives and property in the Loess Plateau area.Therefore,its research is of great significance for protecting the resources and environment of the Loess Plateau,guiding the soil and water conservation work in the Loess Plateau,and revealing the development law of the Loess landform.However,the traditional field survey method is mostly used in the research of loess sinkholes,which has high costs and low efficiency.the existing framework of the loess sinkhole extraction from remote sensing still needs further perfection,and erosion in the loess sinkhole still lacks a better evaluation method.In view of the above problems,based on high-resolution remote sensing images and digital elevation model data,this thesis adopts the combination of object-oriented and deep learning algorithm method to realize the automatic extraction of loess sinkholes and selects the complete watershed to analyze the characteristics of loess sinkholes.Combined with previous surveys,the characteristic factors affecting the development of loess sinkholes are selected from multi-source data,and the prediction model of loess sinkhole erosion-prone areas is constructed to achieve evaluation of loess sinkholes erosion in the target area.The main conclusions are as follows:(1)In this thesis,the optimal segmentation scale of multi-scale segmentation is determined by combining the global Moran’s index with the gray level co-occurrence matrix entropy,and various features of image objects are effectively extracted.Two loess sinkhole extraction models of deep neural network and convolutional neural network are constructed and applied to four verification sample areas.By comparison,the average F1 scores of the two models are more than 83%,and the average F1 score of the convolutional neural network extraction model is 85.99%.The comprehensive extraction performance and generalization ability are good,and the loess sinkholes in the target area can be quickly extracted.(2)In this thesis,Based on the analysis of the characteristics of 877 loess sinkholes in the Maliangou watershed,it is found that the distribution of sinkholes in the watershed is mainly located near the first and second-level river networks,and it is aggregated in the gully area.The side shows that the development of sinkholes is closely related to the gully erosion process.Some sinkholes are distributed on the edge of the loess tableland and the loess mound;The top topography of the loess tableland and mound is relatively flat,and the development of sinkholes is less;In terms of morphology,most of the sinkholes shapes in the basin can be well fitted by ellipses,and the diameter length is concentrated between 0 and 5 meters.(3)In this thesis,by comprehensively considering the internal and external factors of loess sinkhole development,12 feature factors are extracted from multi-source data and screened and combined to construct a variety of prediction models for sinkhole erosion-prone areas.It is verified that the accuracy,precision,recall,rate,and F1 score of the Light GBM model trained based on the combination of elevation,aspect,plane curvature,slope,vegetation coverage,soil erodibility,slope length and soil and water conservation factors are all over 82%,and the prediction results are basically consistent with the development characteristics of loess sinkholes,which can effectively evaluate the erosion of sinkholes in the target area. |
PDF Full Text Request