Mapping Of Himalaya Leucogranites Based On Metric Learning

Formed as a result of the collision of the Indian and Asian subcontinents,the Himalayan orogenic belt has attracted growing attention owing to its complex geological structures,frequent magmatic activities,and abundant mineral resources.Previous studies have indicated that the Himalayan leucogranites are closely associated with rare metal mineralization,such as Be,Nb,Ta,Sn,W,U,and Li.These findings imply that the Himalayan leucogranite belts are expected to be important metallogenic belts for rare metals in China.Rare metals play a considerable role in the development of new materials and energy,making them key mineral resources for global competition.Thus,finding a way to effectively map the distribution of Himalayan leucogranite and carry out theoretical and applied research are critical for the exploration of rare polymetallic mineral resources in the Himalaya region.However,significant uncertainty surrounds the distribution of Himalayan leucogranites owing to the poor environment with regard to the lack of more precise information on the geological,geochemical and geophysical data,which restricts further mineral exploration of the rare metal polymetallic deposits in this area.Fortunately,the sparse air and vegetation cover provide favorable conditions for the application of remote sensing technology.Remote sensing technology has long been recognized the significance in geological works,which greatly promoted mineral exploration in a cost-effective manner,especially in the Himalayan orogenic belt with poor natural environment.In a word,this paper aims to establish lithological mapping models for accurately delineating the spatial distribution of leucogranites with limited geological data,thereby provides technical support for further exploration of rare metal deposits in the Himalayan orogeny belt.Several challenges exist in relation to either recognition methods or data processing,which can be summarized as the following:(1)limited spatial sampling of geochemistry in the study area(sampling density of 1 per 4 and 16 km~2in the study area);(2)implementation of assumptions of data distribution for traditional unsupervised methods(e.g.,band ratio and K-means),whereas supervised methods(e.g.,random forest and k-nearest neighbors)are incompetent in establishing accurate models for identifying highly similar lithologic units with limited training samples.To solve the problems mentioned above,this paper carries out several research works from the sides of data processing and identification methods,on the basis of two study areas with different scales.Regarding the first issue,multi-source data fusion was first employed to provides more information for lithological mapping.In relation to the second issue,this paper develops three supervised metric learning approach,intend to improve the lithological mapping models,and further improve the recognition of leucogranite.Overall,the main works and conclusions are summarized as follows:(1)Framework for leucogranite mappingA framework of leucogranite mapping based on geochemical data,Aster and Sentinel-2 remote sensing images,and metric learning models was established on two study areas with regional and local scale,respectively.Namely,the regional geochemical data is used to map the high potential area of leucogranites.Then,large scale lithologic mapping is carried out based on the fusion data in the selected study area,which intend to draw the accurate distribution area of leucogranites.(2)Multi-source data fusionMulti-source geoscience data(e.g.,geology,geochemistry,geophysics,and remote sensing)provide various types of information for mineral exploration.These data reflect the physical and chemical characteristics of geological units in different ways.A multivariable analysis fusion technology was first employed to merge lower-resolution ASTER data and higher-resolution bands in Sentinel-2A data.This process improved the spatial resolution while maintaining the original spectral characteristics of the ASTER image.Besides the remote sensing images fusion,this paper combines geochemical data at a scale of 1:200,000 with 30m resolution ASTER image according to their correlation.The fused data complements several areas without geochemical samples,moreover,retains both the geochemical distribution information and clear spatial details and texture of the ASTER image.Multi source data fusion helps reduce possible uncertainties and maximizes the utility of various types of data by providing more useful and complete information.(3)Lithological mapping based on metric learningLithological mapping can be performed to determine the target intrusions by learning the distance between the unknown and labeled data.Metric learning,defined as the measure of similarity between samples,describes a more optimal distance approach based on the Mahalanobis covariance distance rather than the Euclidean distance.Metric learning enables to evaluate the similarity of geological data by obtaining accurate spatial description,which is significant for improving the performance of lithological mapping model by converting the original data into more separable spaces.As a result,a weighted metric learning-based approach that combines discriminative local ensemble learning with the SVM classifier was successfully developed to map leucogranites in the Himalayas based on the regional geochemical dataset.To further improve the recognition of lithologic boundaries,this paper established two metric learning-based lithological mapping models in Cuonadong dome,the northern part of the Himalayan orogeny belt,namely maximum margin metric learning and random forest metric learning.To measure the performance of these two models,seven types of lithologic units were selected for lithologic mapping based on the fusion data.These lithological units were effectively discriminated with a maximum 87.8%classification accuracy of leucogranite.This superior performance illustrated the effectiveness of metric learning in improving separability and efficiency of geological dataset.In this paper,several metric learning-based methods were introduced to map Himalayan leucogranites in the Himalayan orogen,thereby providing alternative ways of identifying favorable intrusions based on geochemical exploration data and remote sensing images in poor environments that have undergone limited geological research.The contributions are mainly reflected on the sides of data processing and identification methods.In short,this paper first provides abundant and accurate data for lithological mapping through data fusion,which solves the problem of establishing prediction model due to the lack of enough geological data in the study areas.Facing highly similar geochemical samples and highly correlated and redundant remote sensing data brought by data fusion,this paper developed three metric learning-based lithological identification models to map Himalayan leucogranites in the selected local and regional study areas.Visual and quantitative evaluation indexes confirmed that metric learning can effectively reduce the complexity of data and improve the recognition accuracy of leucogranites.More importantly,this paper sought to establish an important scientific basis for further exploration for rare metal deposits in the Himalayan orogeny belt,and a way of thinking for detecting geological features under harsh natural conditions.