Technology of classifying lithology of typical karst areas based on machine learning:Taking the Pingguo area,Guangxi as an example
Karst is strongly developed in Southwest China,and the geological hazards specific to karst areas may cause serious damage to the local construction facilities;therefore,the rapid identification of carbonate rocks is of great significance for the planning of infrastructure construction such as electric power,transportation,etc.At present,the classification of lithology through remote sensing is still one of the most efficient ways.In this study,on the basis of visible to short-wave infrared multispectral information,texture information such as the second order moments of entropy and angular,and topographic features such as curvature and slope,we proposed a method to classify carbonate rocks,clasolite rocks,Quaternary sediments and water bodies by remote sensing in the Pingguo area to realize automatic extraction of carbonate rocks.This method presents its advantages of obtaining data sources with high accuracy in simple realization path in terms of lithology classification of typical karst areas,and this study can provide a new idea for rapid lithology classification in karst areas.Located in the southwest of Guangxi with subtropical monsoon climate,the Pingguo area is one of the most developed karst areas in China,in which mainly develops limestone,dolomite and their interbedded and interlayered layers of the Triassic Beisi Group and Luolou Group.Karst landforms are distributed in the north,central and southwest of the Pingguo area,and are dominated by the karst peaks and depressions,significantly different from the morphological characteristics of clasolite rocks in the same area.The process of lithology classification and carbonate rock identification in this study mainly included the following steps:data source selection and processing,texture feature extraction,terrain feature extraction,sample sketching and training,and final mapping.The data was mainly from Landsat 8 OLI and AW3D 30 DSM available for the public.The 8-channel data was obtained,based on the synthesis of five multispectral bands including BLUE,GREEN,RED,NIR,SWIR bands,texture features of second-order moments of entropy and angular with Principal Component Analysis,and topographic features such as curvature and slope.Then the maximum likelihood classification was used to carry out the lithology classification of the study area,and the distribution of carbonate rocks was finally obtained.The results show that the overall accuracy of the method in the study area is over 90％,and the kappa coefficient is more than 0.85,which proves that the classification model is effective and practical for the extraction of carbonate rocks in the study area.In addition,results of accuracy evaluation show that the accuracy is high for classification of carbonate rocks,with an accuracy more than 94％,but the accuracy of classifying water bodies and Quaternary is relatively poor.The main reason for this difference is that the Quaternary sediments are scattered in the carbonate and clasolite areas,and part of Quaternary sediments may exist in the area covered with vegetation which is easy to be judged as the area of carbonate or clasolite rocks.The main reason for the slightly lower accuracy in classification of water bodies is that some water bodies and the areas shaded by carbonate rocks are easily confused,and there are also cases where carbonate rocks are misclassified as water bodies.Based on multi-source remote sensing data,we conducted the maximum likelihood estimation with the consideration of spectral,topographic,textural and other multi-factors in order to develop a fast and accurate automatic extraction method for carbonate rocks in typical karst areas.In addition,we used this method to automatically classify carbonate rocks,clasolite rocks,Quaternary sediments and water bodies in the Pingguo area and evaluated the classification accuracy.We drew the conclusions as follows.(1)Based on data of Landsat and AW3D 30 DSM,we adopted multi-factor synergistic analysis of spectral information,textural features and topographic features of typical carbonate areas,and carried out automatic extraction of carbonate rocks,clasolite rocks,etc.in the Pingguo area,a typical karst area.The overall classification accuracy was greater than 93％.(2)There was a great difference in the surface morphology of carbonate rocks and clasolite rocks in the study area,which is one of the important factors for distinguishing these two types of rocks.Meanwhile,the slope has a significant effect on the accuracy of carbonate rock extraction,so it is suggested that a slope should be used as one of the factors in the classification in the process of comprehensive extraction of carbonate rocks.(3)Lithology is difficult to be automatically extracted with multi-spectral remote sensing information,due to the complex multi-solution of extraction and the spatial heterogeneity of lithology.Most of the current studies differentiated lithology through the inversion of mineral compositions from the perspective of mineralization and alteration.With the gradual development and improvement of machine learning and deep learning,the comprehensive extraction method integrated with spectral bands,texture and morphology will provide a new idea for the significant improvement of the accuracy of lithology extraction,which may be one of the important directions for future research on lithology extraction.
国家科技期刊平台
NSTL
万方数据
