Building The Convolutional Neural Network And Lithological Mapping Based On Geochemical Exploration Data

Traditional geological mapping,primarily based on limited outcrops and investigations,may be greatly affected by the experience and professional knowledge of the mapping geologists,and has certain uncertainties.Geochemical survey data contain rich geological information such as lithology and structure,which can achieve the purpose of dividing lithology through shallow overburden layer,and then provide more effective clues for geological mapping.Therefore,it is of great scientific significance for basic geological research and regional geological survey to accurately identify the distribution of underground lithologic units by digging deep spatial information in geochemical survey data.Conventional geological mapping methods(such as similarity measurement methods,shallow machine learning,etc.)are pixelwise-based modes which usually integrate multiple information at a single pixel,do not adequately consider the spatial associations among neighboring pixels and ignore the spatial characteristics of the neighboring data.The classification model or combination algorithm in deep learning algorithm has strong feature extraction ability for nonlinear data.For example,The hierarchical algorithm using multi-level nonlinear transformation in convolutional neural network can fully consider the spatial characteristics of data,and can be used as an auxiliary means with potential advantages to improve the efficiency and accuracy of lithological mapping.The West Qinling Orogen is an important gold polymetallic metallogenic belt in China.It is of great significance to know the lithology distribution in this area for further mineral resource exploration.The study area,Daqiao gold district,is located in the eastern part of the West Qinling Orogen,China.Based on the study of the basic geological characteristics of the study area,the sliding window technology was used to construct training samples with spatial characteristics.According to the differences in the distribution patterns of geochemical elements in different lithologic units,the deep learning algorithm was introduced to deeply explore the potential spatial information features.The lithology identification model of convolutional neural network based on Alex Net is established to efficiently delineate the spatial distribution range of seven lithological units,which provides technical support for the geological survey of Daqiao gold district in Gansu Province.In order to solve the convolution neural network parameter optimization problem,this thesis introduces two automatic hyperparameter optimization algorithms,analyzes and discusses the performance of the network model from the operating mechanisms of deep learning such as network hyperparameters and network structure.By comparing the experimental results of grid search and random search,and combining them with manual debugging,the optimal hyperparameter configuration is obtained in the current lithology identification task,and the time of hyperparameter optimization is greatly reduced.The experimental results show that the convolutional neural network model based on geochemical survey data can effectively identify seven lithological units in the study area.The geological mapping roughly match with the original geological map,and the overall classification accuracy can reach 90%.It not only proves that geochemical survey data have the potential to identify lithologic units,but also shows that convolutional neural network can effectively extract potential geochemical features from geochemical survey data.In order to reduce the uncertainty caused by the insufficient data mining on the model test results,this thesis discusses the influence of the selection of training samples at different locations on the performance of the convolutional neural network model,which further confirms the application value of the convolutional neural network method in intelligent lithology mapping based on geochemical survey data.