Research On Lithology Recognition Method Based On Hyperspectral And Machine Learning

Lithology identification is an important factor in the field of petroleum exploration and development.It is of great significance to establish a lithology identification model with high efficiency and good performance.However,lithology identification is confronted with many difficulties:complex geological lithology information,many interference factors in rock sample information collection,and strong expertise in rock sample information analysis.The traditional lithology identification methods still have low efficiency,low accuracy and poor stability.Hyperspectral detection technology has high efficiency,fast speed,simple operation,and no pollution,widely used in the environment,geology,chemical industry,and other fields.The key to applying hyperspectral detection technology to lithology analysis is to find a suitable data processing method and construct an accurate identification model.However,the local differences of conventional data processing methods are not easy to enlarge and feature extraction ability is insufficient.The traditional modeling method has poor applicability and high professional requirements.Therefore,a rapid lithology identification method is proposed by combining hyperspectral analysis and a machine learning algorithm.On this basis,data preprocessing and dimension reduction methods are added to further improve the performance of the model from different aspects.The main research contents and conclusions are as follows:（1）Taking the Bahe area of Lantian County as the first study area,4 types of sandstones with different depths of tight sandstone reservoir in the same cored well were collected,and the Dagang area of Binhai New area as the second study area,4 types of typical sedimentary rocks with different exploration Wells were collected,and the hyperspectral data of 4 types of cores including sandstone,mudstone,limestone and shale in the study area were collected,and the corresponding reflected hyperspectral data were obtained.The reflection hyperspectral characteristics of four kinds of rock samples and the influence mechanism of corresponding characteristics are obtained.Among them,the absorption valley characteristic before 1400nm is related to the energy level transition of electrons of Fe²⁺and Fe³⁺ions in the rock composition.The characteristics of the absorption valley after 1400nm are related to the bending or stretching vibration transitions of CO₃^2-and OH^-groups in the rock samples,as well as the water molecules in the rock samples.Among them,the more obvious characteristic peak near 2200nm is affected by the lattice vibration of Al-OH and Mg-OH in the composition of rock samples.According to the reflection hyperspectral characteristics and influence mechanism of the above samples,the differences in reflection hyperspectral characteristics of four typical sedimentary rocks are obtained.（2）Given the problems of noise interference and the same spectrum of foreign bodies in rock hyperspectral data,the reflection hyperspectral curves of rock samples in the band range of 1800-2400nm were selected.Three data preprocessing methods,Savitzky-Golay convolution smoothing（SG）,multiple scattering corrections（MSC）,and standard normal variable transformation（SNV）,were used to eliminate interference and further improve the accuracy of model recognition.Among the established models of extreme learning machine（ELM）,probabilistic neural network（PNN）,and support vector machine（SVM）,the hyperspectral data preprocessing method using SNV has the best effect,which can effectively eliminate background noise and amplify local hyperspectral differences of similar rock samples.The accuracy of the external test set model was improved by 12.50%,11.25%,and12.50%,respectively.（3）Given the low efficiency and low accuracy of rock hyperspectral detection,three feature extraction methods,namely principal component analysis（PCA）,competitive adaptive reweighted sampling method（CARS）,and successive projections algorithm（SPA）,were used to extract the hyperspectral variables of rock samples.The extracted feature variables were fed into the SVM external test set model based on the full spectrum.The performance improvement effects of the three feature extraction methods were ranked in descending order:SPA,CARS,PCA.The accuracy improved by 8.33%,10.42%,and 12.50%,respectively.The SPA model abandons a large number of redundant information,effectively extracts the key information in the data,and improves the efficiency and accuracy of the model.（4）Aiming at the problem of few training samples of hyperspectral data of rock samples,based on the data processing method,an SVM hyperspectral data recognition model of rock samples based on the full spectrum was established,and ELM and PNN models were constructed for performance comparison.Among them,the SVM model is more suitable for processing complex data and small sample data,the accuracy of the training set and test set are96.55%and 100%,and the Kappa coefficient is 0.9941 and 0.9792,respectively.The accuracy of ELM and PNN model test sets was 95.83%,and the Kappa coefficients were 0.9199 and0.9255,respectively.So far,the optimal combination model for lithology identification is SNV-SPA-SVM,the number of extracted characteristic variables is 9,and the accuracy of the test set is 100%.The research methods and conclusions show that the successful combination of hyperspectral detection technology and machine learning algorithm is simple and easy to use,which can quickly and accurately identify different kinds of rock samples,and has broad application prospects in oil and gas field exploration and development.