深度残差神经网络频率域下蚀变岩石光谱分类模型研究  

Research on Altered Rock Spectrum Classification Model in Frequency Domain of Deep Residual Neural Network

作　　者：满文婧 陈文龙 徐元进[1] MAN Wen-jing;CHEN Wen-long;XU Yuan-jin(China University of Geosciences,Wuhan 430074;CCCC Highway Consultants Co.,Ltd.,Beijing 100010,China)

机构地区：[1]中国地质大学(武汉),湖北武汉430074 [2]中交公路规划设计院有限公司,北京100010

出　　处：《地理与地理信息科学》2021年第3期29-34,F0002,共7页Geography and Geo-Information Science

基　　金：国家自然科学基金面上项目(41872252)。

摘　　要：以往的蚀变岩石光谱分类方法包括机器学习模型和浅层神经网络模型,且均使用光谱波形特征作为模型分类依据。小波变换能够将光谱数据转换为频谱图进行奇异性探测,显示光谱瞬态突变信息。该文使用5种不同蚀变类型的石英闪长玢岩光谱数据,通过Symlets小波变换将光谱数据转换为频谱图,作为深度卷积分类模型的基础数据,再利用50层深度残差网络(ResNet50)模型对数据进行分类,并选取准确率、损失值、召回率、精确率以及F 1参数对结果进行评价。通过与传统的机器学习方法及其他方法进行比较,证明该模型的准确率和F 1参数均优于对比模型,测试集分类准确率达到99.67%,表明该模型对蚀变岩石光谱数据分类的适用性较强,且具有较好的鲁棒性和泛化性。As an important branch of geological science,altered rock identification was based on machine learning or shallow learning network in previous studies,which used spectral characteristics as the basis of model classification.Continuous wavelet transform can perform singularity detection by converting spectral data to frequency spectrum to display transient mutation information and local time-frequency characteristics.This paper puts forward an altered rock classification model.In this model,the spectral data are converted to the frequency domain by wavelet transform,and the converted spectrum is used as the input of the deep convolution classification model to classify the altered rocks.In order to verify the feasibility of the model,symlets wavelet transform was used in the experiment to convert the spectral data of 5 different alteration types of quartz diorite porphyrite into time-spectrum diagram,and then ResNet50 model was used to classify the altered rocks.The classification results were compared with those of PCA+SVM model,PLSR model,RF model,AlexNet model,ResNet101 model,ResNet152 model and 1D-ResNet50 model.The model classification results were evaluated by selecting accuracy,loss value,recall rate,precision rate and F 1 score.From the experimental results,it can be concluded that the accuracy and F1 score of the proposed model in this paper are superior to the comparison models.And the classification accuracy of the test set reaches 99.67%.The results show that the model has good applicability,robustness and generalization to the classification of spectral data of altered rocks.

关 键 词：蚀变岩石光谱 小波时频变换 深度残差网络 

分 类 号：P585[天文地球—岩石学]