Research On Remote Sensing Image Prospecting Prediction And Mapping Method Based On Multi-scale Analysis And Machine Learning

With the rapid development of remote sensing data acquisition technology,the quantitative and intelligent processing technology of remote sensing data needs to be improved.Remote sensing image based exploration,deposit location and mapping technology can significantly improve the efficiency and quality of mineral resources exploration and mapping.In recent years,great progress has been made in prospecting and mapping based on remote sensing technology,but there is still a gap between the application effect and the actual needs of geological survey.There are several difficulties in prospecting prediction and mapping based on remote sensing images:1)Pixel based matched filtering,band ratio,principal component analysis and other technologies can not eliminate the influence of climate very well,light and other factors in the process of image acquisition,nor can they make good use of mineral cluster characteristics,and the extracted alteration information has more "salt and pepper" noise;2)For the extraction of remote sensing geological information in areas with many vegetation cover,many interference information and weak mineralization clues,the mask method will lose original information of the image,which can not guarantee the accuracy of the extraction results;3)The single scale analysis method can not effectively extract the multi-resolution characteristics of mineral anomaly distribution,and can not accurately describe the enrichment and dilution of minerals;4)The intermittence and multi-stage of hydrothermal alteration lead to the superposition of spectral information and the texture uncertainty caused by lithologic weathering,which lead to the decrease of the accuracy of lithologic identification.These problems become the bottleneck of large-scale remote sensing geological application.In this paper,with the aid of field geological survey and thin section identification,multi-scale analysis method,object-oriented extraction technology,machine learning and depth feature decomposition as the main line,the regional feature and depth feature extraction method,lithology intelligent classification and mapping method of altered minerals are studied.The main research contents and achievements are as follows:(1)In this paper,a remote sensing mineralization alteration information extraction method based on principal component analysis,multi-scale segmentation and support vector machine is proposed.In this method,the diagnostic bands of mineralization and alteration information in ASTER image are selected for principal component analysis;the singularity and self similarity of minerals are described by using multifractal theory to obtain multi-scale texture image;most irrelevant data are filtered by using local features,and the target mineral category is located by using support vector machine vector approximation method;the sequence minimum optimization algorithm can improve the efficiency of solution.The experimental results show that the information of mineralization and alteration extracted by this method has a good correlation with metallogenic belts,known ore spots and metallogenic characteristics of different geological backgrounds.(2)On the basis of principal component analysis of altered mineral feature vector,a method of extracting alteration information combining wavelet packet transform and random forest is proposed.In this method,wavelet packet transform is used to extract the time-frequency localization and multi-scale detail features of the image,and the cost function is used to optimize the wavelet packet tree to obtain the optimal representation of the high and low frequency information of altered minerals.The important features are screened through the interference feature mechanism,and the voting classification is completed by using random forest.The experimental results show that this method can make full use of the energy characteristics of mineral spectrum and reduce the noise interference of mineral components when extracting the alteration information of iron staining,Al oh and Mg OH groups.(3)A method of deposit location based on multi-scale convolution neural network feature decomposition is proposed.Based on the analysis of the organic matter characteristics of metal deposits,this method makes full use of the image forms such as color,shape and texture displayed by the image to construct the deep semantic information classification space;uses fuzzy mathematics theory,element multiplication algorithm intersection,logical superposition analysis to extract the ore controlling factors of the image;combines with multi-source data such as geophysical and geochemical exploration to construct the remote sensing geological prospecting model.The experimental results show that this method provides a reliable basis for peripheral exploration and deposit location in geological exploration.(4)The method of ratio operation,multi-scale segmentation and random forest is proposed to extract metamorphic minerals.The core idea is to use the variogram function in geostatistics to describe the global and local spatial structure variability of minerals.The multi-scale texture features and spectral features are combined by vector superposition method,and the random forest is used to search the mineral distribution zone.The experimental results show that this method can well describe the randomness of geochemical element distribution and the local enrichment and dilution of geochemical elements in rocks and other media,and the extraction results are stable.(5)A mapping method for automatic classification of lithology and cross validation of main and typical rock forming minerals is proposed.The core idea is to use the image spectral features and multi-scale texture features obtained by wavelet transform to construct the classification feature space,to carry out 10 times of high-dimensional and non normal distribution of lithology classification,to use the voting method to avoid the dynamic changes of lithology due to the spatial variability of samples,and to optimize the classification results of lithology units;to use the bee colony algorithm which can avoid local optimization to search the parameters of support vector machine,the index of effective extraction of muscovite,biotite,calcite,amphibole and other minerals is constructed.Field work shows that this method can intelligently identify most of the lithology of the image,and the correlation coefficient between the mapping results and the field survey results is 0.7.