Research On Lithology Classification Of Multi-source Remote Sensing Data Supported By Machine Learning

With the rapid development of remote sensing technology,the spectrum range of remote sensing detection is expanding,the resolution is improving,and the available data sources are increasing.Because different lithologic units have different display characteristics on different types of data sources,how to integrate the advantages of multi-source remote sensing data and apply them to Remote Sensing Lithologic mapping based on artificial intelligence algorithms and models is a major issue,and there are still many aspects to be explored and studied.Combining remote sensing and GIS technology,three kinds of multi-source remote sensing data,GF-1 satellite data,Landsat-8 OLI satellite data and Sentinel-1A radar data,are selected.On the basis of image preprocessing,interference information weakening and multi-source data collaboration,Construction of local Lanczos double-diagonalization limit learning machine model and support vector machine model are used to classify lithology Dejiang area,Guizhou Province,respectively.This study provides some technical references for other multi-source remote sensing data cooperative processing,and also has a certain reference role for other artificial intelligence algorithms to introduce remote sensing image classification field.The main achievements are as follows:(1)Using the method of "suppression-mask-forced invariance-histogram equalization" to suppress vegetation information,the Landsat-8 OLI data processing experiment in Dejiang area of Guizhou Province shows that this method can significantly suppress the vegetation information in the original image.It is proved that this method is a general optical remote sensing image vegetation information suppression method which does not need prior knowledge and field spectral data,and is suitable for areas with complex topography,more shadows and medium vegetation coverage.(2)The texture information of high-resolution data,spectral information of medium-resolution data and stereo structure information of radar data are sequentially processed by pixel collaboration and feature collaboration.The results show that the texture principal component transform fusion method can not only eliminate redundant data and concentrate variance information,but also preserve the spatial structure information of objects.The Gram-Schmidt transform fusion method has strong comprehensive performance,which makes the fusion image spectral fidelity very good,and the boundary of various objects more obvious,and the discrimination is greater.(3)The LBD-ELM model and SVM model are constructed to classify the lithology of the study area by using multi-source collaborative data,supplemented by useful feature information-spectral and topographic data.The results show that the overall classification accuracy of LBD-ELM model is 88.12%,Kappa coefficient is 0.8534;the overall classification accuracy of widely used SVM model is 86.58%,Kappa coefficient is 0.8268;the classification results of LBD-ELM model have better patch integrity and fewer errors and omissions.Therefore,as a whole,LBD-ELM model can meet the needs of remote sensing lithology mapping to a greater extent.