Carbonate rock lithology identification method based on machine learning

Abstract

The invention provides a carbonate rock lithology identification method based on machine learning, and the method comprises the steps: obtaining the hyperspectral data of a carbonate rock sample through an ASD spectrometer, and carrying out the preprocessing of the hyperspectral data; a mixed feature selection method is adopted, firstly, feature wavebands in the hyperspectral data are selected, then a random forest method is adopted for feature importance sorting, and features with high importance are selected; constructing a carbonate rock lithology identification model based on the selected features, wherein model parameters are determined by a machine learning algorithm; and predicting the mineral content in the carbonate rock according to the lithology identification model so as to identify the lithology of the carbonate rock. According to the method, on the basis of carbonate rock hyperspectral data, feature selection is carried out by adopting a mixed feature selection method after preprocessing, a lithology identification model is established on the basis, model parameters are determined through a machine learning mode, mineral component content is predicted, and therefore carbonate rock lithology identification and classification are achieved. And the identification precision and efficiency are greatly improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of oil and gas exploration and development and geological detection, and relates to a carbonate rock lithology identification method based on machine learning, and in particular to a carbonate rock lithology identification method based on spectral feature analysis and machine learning. Background Art

[0002] Carbonate rock is one of the main components in the strata, and is also an important oil and gas reservoir rock and source rock. Before the exploration and development of oil and gas, the structural properties of the underground reservoir need to be tested. In addition to oil and gas reservoirs, there are also carbonate reservoirs in the underground reservoirs. Their composition and lithology distribution have an important influence on the storage and distribution of oil and gas. Therefore, the identification of carbonate rocks has important guiding significance for oil and gas geology and oil and gas exploration and development.

[0003] Traditionally, the identification of carbonate rock lithology is mainly done by experienced geologists who distinguish between limestone and dolomite by observing samples indoors or observing the color, texture, structure and other characteristics of rocks in the field. In some cases, chemical reagents are needed for identification. For example, the method of identifying carbonate rock lithology at a construction site is to manually select a certain amount of carbonate rock samples, grind them, put them in a closed container and react with dilute hydrochloric acid, and determine the content of calcium carbonate and magnesium carbonate in the samples based on the pressure of the generated gas, and then describe the lithology of the samples. The above method is not only inefficient, but also greatly affected by subjective human factors during the identification process, which is not conducive to the rapid and accurate identification of carbonate rock lithology, especially it is difficult to distinguish between many lithologies such as limestone, dolomite-bearing limestone, lime-bearing dolomite and dolomite in carbonate rocks.

[0004] An ore body or ore deposit is a rich body of ore-forming elements, which are often accompanied by alteration. Since altered minerals have their own spectral characteristics, minerals can be identified by the spectral characteristics of rocks and minerals. CN 113049521A discloses a method, device, equipment, and storage medium for identifying carbonate rocks. The method includes: obtaining a rock to be identified; testing the rock to be identified to obtain its thermal infrared spectrum data; calculating the relative depth of the reflection peak of the thermal infrared spectrum data in the first preset band interval to obtain a first relative depth value; calculating the relative depth of the reflection peak of the thermal infrared spectrum data in the second preset band interval to obtain a second relative depth value; detecting whether a preset condition is met; the preset condition includes: the first relative depth value is greater than the first preset threshold value, or the second relative depth value is greater than the second preset threshold value; if the preset condition is met, it is determined that the rock to be identified is carbonate rock. This method can only identify carbonate rocks and other types of minerals, but cannot effectively identify different lithologies of carbonate rocks, and has a narrow scope of application.

[0005] CN 112240875A discloses a carbonate rock lithology identification method and device based on ASD spectrum, the method comprising: using an ASD spectrometer to obtain spectral data of carbonate rock samples; preprocessing the spectral data; extracting local waveform characteristic values ​​in the spectral data; establishing a carbonate rock lithology identification model according to the local waveform characteristic values ​​in the spectral data; and identifying limestone and dolomite using the carbonate rock lithology identification model. Although the method improves the recognition efficiency and accuracy of carbonate rock lithology compared with the manual recognition method, the method of establishing the recognition model still relies on the waveform characteristic values ​​in the spectrum, the feature selection method is relatively single, and a reference model applicable to more samples is not established. It can only identify dolomite and limestone, but cannot identify different lithologies that are further finely divided.

[0006] In summary, for the method of carbonate rock lithology identification, it is necessary to combine more samples and feature selection methods on the basis of spectral data, and establish an identification model with a wider range of applications and higher identification accuracy through machine learning, so that it can further identify the different lithologies of dolomite, lime-bearing dolomite, dolomite-bearing limestone and limestone in carbonate rocks. Summary of the invention

[0007] In view of the problems existing in the prior art, the purpose of the present invention is to provide a carbonate rock lithology identification method based on machine learning. The method is based on the hyperspectral data of carbonate rock samples, and after preprocessing, a hybrid feature selection method is used for feature selection. Then, a lithology identification model is constructed with the selected features through a machine learning method to identify and classify the lithology of carbonate rock samples, thereby effectively improving the efficiency and accuracy of lithology identification and providing a basis for geological exploration and oil and gas development.

[0008] To achieve this object, the present invention adopts the following technical solutions:

[0009] The present invention provides a carbonate rock lithology identification method based on machine learning, and the lithology identification method comprises the following steps:

[0010] (1) Using ASD spectrometer to obtain hyperspectral data of carbonate rock samples and preprocessing the hyperspectral data;

[0011] (2) Using a hybrid feature selection method, first select the mineral characteristic bands in the hyperspectral data, then use the random forest feature selection method to sort the features according to their importance and select the features with high importance;

[0012] (3) constructing a carbonate rock lithology identification model based on the features selected in step (2), wherein model parameters are determined based on a machine learning algorithm;

[0013] (4) Predicting the mineral content in the carbonate rock according to the carbonate rock lithology identification model in step (3), thereby identifying the lithology of the carbonate rock.

[0014] In the present invention, the identification of carbonate rock lithology needs to be carried out according to the characteristics of carbonate rock. The commonly used methods currently include using spectral features to identify minerals. The present invention uses an ASD spectrometer to obtain high-resolution reflectance spectral data of carbonate rock. By preprocessing the hyperspectral data, the absorption and reflection characteristics of the spectral curve are effectively highlighted, noise is suppressed, and the classification and identification of carbonate rock are facilitated; by selecting features in the spectral data, especially selecting the characteristic bands of carbonate rock, and selecting features with high importance through the random forest feature selection method, and then establishing a lithology identification model based on this, and determining the model parameters through machine learning, and predicting the content of mineral components through the lithology identification model, the identification and classification of carbonate rock lithology are obtained. Compared with the manual identification method, the method has greatly improved work efficiency and significantly improved the accuracy of lithology identification, which can provide a basis for geological detection and oil and gas exploration and development of oil and gas reservoirs, and has broad application prospects.

[0015] The following are preferred technical solutions of the present invention, but are not intended to be limitations of the technical solutions provided by the present invention. Through the following technical solutions, the technical objectives and beneficial effects of the present invention can be better achieved and realized.

[0016] As a preferred technical solution of the present invention, the ASD spectrometer in step (1) is a portable ground object spectrometer.

[0017] Preferably, the hyperspectral data in step (1) is remote sensing imaging spectral data with a spectral resolution of nanometer order.

[0018] Preferably, the hyperspectral data in step (1) is a reflectance spectrum in the band range of 350 to 2500 nm, wherein the wavelength can be selected from 350 nm, 500 nm, 1000 nm, 1500 nm, 2000 nm, 2100 nm, 2200 nm, 2300 nm, 2400 nm or 2500 nm, etc.

[0019] Preferably, the acquisition of the hyperspectral data in step (1) needs to be corrected by the reflectivity of a white board.

[0020] In the present invention, the spectral data collected by the ASD spectrometer is calibrated with reference to the whiteboard data. The whiteboard reflectance data is obtained in advance through laboratory calibration. The specific formula is as follows: ground object reflectance = (ground object DN value / whiteboard DN value) × whiteboard reflectance.

[0021] As a preferred technical solution of the present invention, the preprocessing in step (1) includes outlier detection and elimination and decontinuum processing.

[0022] Preferably, the outlier detection and elimination is to eliminate sample spectra with abnormal hyperspectral data from the hyperspectral data of a large number of samples.

[0023] Preferably, the detection of the abnormal points adopts the upper and lower cutoff point method in statistics.

[0024] Preferably, the selection of the upper and lower cutoff points is calculated using the quartile method, wherein the quartiles include the first quartile Q1, the median Q2 and the third quartile Q3.

[0025] In the present invention, the presence of outliers has a great impact on the accuracy and generalization ability of the machine learning model. In order to improve the accuracy of the model, it is necessary to detect and delete the outliers in the data in advance. There are many ways to detect outliers. The upper and lower cutoff point method used in the present invention, for each band, statistical analysis is performed based on the values ​​of all samples in the band, and the first quartile Q1, median Q2 and third quartile Q3 are extracted, and the value of the upper cutoff point is Q3+1.5×(Q3-Q1), and the value of the lower cutoff point is Q1-1.5×(Q3-Q1); if the value of the sample in the band is above the upper cutoff point, or below the lower cutoff point, the number of outliers is counted as 1, and the number of outliers for each sample is accumulated. If the number of outliers exceeds 75%, the spectral data of the sample needs to be regarded as abnormal and eliminated.

[0026] As a preferred technical solution of the present invention, the decontinuum processing in the preprocessing includes the generation of a continuum and a normalization processing of the continuum elimination method.

[0027] In the present invention, the purpose of the decontinuum processing is to effectively enhance the absorption and reflection characteristics of the spectral curve, unify the benchmark, and facilitate the comparison of characteristics. The image normalized by the continuum elimination method can effectively suppress noise, highlight the characteristic information of the ground object spectrum, and facilitate the classification and identification of carbonate rocks.

[0028] Preferably, the Clark shell coefficient method is used to generate a continuum of the spectral curve.

[0029] Preferably, the normalization processing by the continuum elimination method comprises: using the actual reflectivity of the sample to remove the reflectivity of the corresponding band on the continuum to obtain a normalized value by the continuum elimination method.

[0030] In the present invention, the continuum is intuitively equivalent to the "shell" of the spectral curve. Since the actual spectral curve is composed of discrete sample points, a continuous broken line segment is used to approximate the envelope of the spectral curve. Specifically, the continuum generation in the present invention adopts the shell coefficient method proposed by Clark, and the specific algorithm is as follows: all the maximum points on the spectral curve, that is, the prominent "peak" value points, are obtained by differentiation, and then the maximum point among all the maximum points is found; the maximum point is used as an endpoint of the continuum, and the slope of the line connecting the point and each maximum value in the direction of increasing wavelength is calculated, and the point with the maximum slope is used as the next endpoint of the envelope, and then the cycle is carried out with this point as the starting point until the last point; then the maximum point is used as an endpoint of the envelope, and a similar calculation is performed in the direction of decreasing wavelength, and the point with the minimum slope is used as the next endpoint, and then the cycle is carried out with this point as the starting point until the starting point on the spectral curve; all endpoints are connected along the direction of increasing wavelength to form a continuum.

[0031] In the present invention, the carbonate rock spectrum range of 350-2500nm is selected for continuum elimination normalization processing. The purpose of this step is to normalize the reflectance to 0-1 and greatly enhance it, so that the characteristic values ​​can be more effectively compared with other spectral curves.

[0032] As a preferred technical solution of the present invention, the hybrid feature selection method in step (2) includes a combination of at least two of a filtering method, an embedded method or an encapsulation method, preferably a combination of a filtering method and an encapsulation method.

[0033] Preferably, the filtering method selects a characteristic interval of 1800 to 2400 nm in the hyperspectral data, and the wavelengths within the characteristic interval include 1800nm, 1850nm, 1900nm, 1950nm, 2000nm, 2050nm, 2100nm, 2150nm, 2200nm, 2250nm, 2300nm, 2350nm or 2400nm, and are divided into 4 characteristic intervals of 1800 to 1920nm, 1920 to 2020nm, 2060 to 2250nm and 2250 to 2400nm.

[0034] Preferably, the characteristic intervals of 1800-1920nm, 1920-2020nm, 2060-2250nm and 2250-2400nm correspond to mineral characteristic bands.

[0035] In the present invention, there may be a certain correlation between the bands in the hyperspectral data, and there is data redundancy, so feature selection is required to reduce the number of features and improve the model accuracy and model operation efficiency. The feature selection methods mainly include filtering, embedded and encapsulated methods, among which the filtering method selects the mineral characteristic bands of 1800-1920nm, 1920-2020nm, 2060-2250nm and 2250-2400nm according to previous studies. The spectral absorption characteristics in this interval are related to the mineral composition and content in the rock, and the interval is divided into 560 characteristic bands.

[0036] As a preferred technical solution of the present invention, the encapsulation method is a feature selection method based on random forests, which ranks the feature bands in the feature interval by importance and selects feature bands with high importance.

[0037] Preferably, the characteristic bands are ranked by importance using a random forest regression model.

[0038] In the present invention, the random forest model is not only widely used in prediction problems, but also has certain applications in feature selection. The main reason is that after fitting the data, the random forest will have an importance measurement for all variables. The larger the value of the measurement, the more important the variable is to the accuracy of the prediction. Therefore, based on the first step of feature selection, the random forest method is used to sort the feature importance and select features with high importance.

[0039] In the present invention, the feature selection method of the random forest adopts the RandomForestRegressor model in sklearn, and the selected model parameters are n_estimators, max_depth and min_samples_leaf, and the adjustment range of the parameters is 100-500, such as 100, 150, 200, 250, 300, 350, 400, 450 or 500, etc., 2-8, such as 2, 3, 4, 5, 6, 7 or 8, etc., and 1-8, such as 1, 2, 3, 4, 5, 6, 7 or 8, etc., and the feature_importance parameter is called to obtain the importance ranking of 500 feature bands. According to the importance ranking, the top 50 feature bands are selected to construct a lithology recognition model, and used as input features for subsequent machine learning.

[0040] Preferably, the highly important features account for 8-12% of the total number of feature bands in the selected feature interval, for example 8%, 9%, 10%, 11% or 12%, etc., but are not limited to the listed values, and other unlisted values ​​within the numerical range are also applicable.

[0041] As a preferred technical solution of the present invention, the features of constructing the recognition model in step (3) are the feature bands with high importance selected in step (2).

[0042] Preferably, during the machine learning in step (3), 70% to 90% of the samples are used as training samples, for example 70%, 75%, 80%, 85% or 90%, etc., but are not limited to the listed values, and other unlisted values ​​within the numerical range are also applicable; the remaining samples are used as test samples.

[0043] As a preferred technical solution of the present invention, the machine learning algorithm in step (3) adopts a support vector machine SVM algorithm, preferably a support vector regression SVR method in the sklearn library.

[0044] Preferably, the support vector regression uses a grid search (GridSearch CV) and cross-validation method to determine the hyperparameter C of the algorithm model.

[0045] Preferably, according to the range of SVR hyperparameters 0.001 to 200, for example 0.001, 0.01, 0.05, 0.1, 0.5, 1, 10, 50, 100, 150 or 200, etc., the step size is adjusted, and the adjusted parameters are used for training, and the parameters with the highest accuracy in the validation set are selected as the hyperparameters of the model. If the hyperparameters are not selected appropriately, underfitting or overfitting problems may occur.

[0046] In the present invention, for the determination of the hyperparameters in the SVR algorithm model, when different hyperparameters are used for training, the first 50 features selected according to importance are input, and the composition and content of the sample are output.

[0047] As a preferred technical solution of the present invention, based on the fact that the carbonate rock mainly consists of calcite and dolomite, the identification model is used to predict the content of calcite and dolomite in each carbonate rock sample.

[0048] Preferably, the measured values ​​of the calcite and dolomite contents in the carbonate rock are obtained by X-ray diffraction (XRD) detection.

[0049] Preferably, the fitting curve and its absolute coefficient R are obtained based on the predicted and measured values ​​of the calcite and dolomite contents in carbonate rocks. 2 and root mean square error RMSE.

[0050] Preferably, the lithology classification of the carbonate rock sample is determined based on the predicted values ​​of the calcite and dolomite contents in the carbonate rock sample.

[0051] Preferably, the lithology classification of the carbonate rock is divided into two categories and four categories, the two categories are dolomite and limestone, and the four categories are dolomite, lime-bearing dolomite, dolomite-bearing limestone and limestone, wherein the classification criteria for the four categories of carbonate rock lithology are shown in Table 1.

[0052] When carbonate rocks are classified into two types, they are simply divided into dolomite and limestone, where the content of the components corresponding to dolomite is equivalent to the sum of the content ranges of dolomite and limestone-bearing dolomite in Table 1, and the content of the components corresponding to limestone is equivalent to the sum of the content ranges of limestone and dolomite-bearing limestone in Table 1.

[0053] Table 1 Description of carbonate rock lithology

[0054]

[0055] As a preferred technical solution of the present invention, the lithology identification method comprises the following steps:

[0056] (1) ASD spectrometer is used to obtain hyperspectral data of carbonate rock samples. The ASD spectrometer is a portable ground feature spectrometer. The hyperspectral data is remote sensing imaging spectral data with a spectral resolution of nanometers, which is a reflectance spectrum in the band range of 350 to 2500 nm. The acquisition of the hyperspectral data needs to be corrected by the reflectance of a whiteboard; and the hyperspectral data is preprocessed, and the preprocessing includes outlier detection and removal and continuum removal. The outlier detection and removal is to remove the sample spectra with abnormal hyperspectral data from the hyperspectral data of many samples. The detection of the outliers adopts the upper and lower cutoff point method in statistics, and the selection of the upper and lower cutoff points is calculated by the quartile method; the continuum removal includes the generation of a continuum and the normalization of the continuum elimination method. The Clark shell coefficient method is used to generate a continuum for the spectral curve. The continuum elimination normalization includes: using the actual reflectance of the sample to remove the reflectance of the corresponding band on the continuum to obtain the normalized value of the continuum elimination method;

[0057] (2) A hybrid feature selection method is used. First, a filtering method is used to select the characteristic interval 1800-2400 nm in the hyperspectral data, and the interval is divided into four characteristic intervals of 1800-1920 nm, 1920-2020 nm, 2060-2250 nm and 2250-2400 nm. The above characteristic intervals correspond to mineral characteristic bands. Then, a random forest feature selection method is used to sort the characteristic bands in the characteristic intervals according to their importance, and the characteristic bands with high importance are selected. The characteristic bands are sorted according to their importance using a random forest regression model.

[0058] (3) constructing a carbonate rock lithology identification model based on the characteristic bands of high importance selected in step (2), wherein the model parameters are determined based on a machine learning algorithm. During the machine learning, 70 to 90% of the samples are used as training samples, and the remaining samples are used as test samples. The machine learning algorithm adopts a support vector regression (SVR) method, and the support vector regression adopts a grid search and cross-validation method to determine a hyperparameter C of the algorithm model. According to the range of the SVR hyperparameter, the step size is adjusted, and the adjusted parameter is used for training, and the parameter with the highest accuracy in the validation set is selected as the hyperparameter of the model;

[0059] (4) Predicting the mineral content in the carbonate rock according to the carbonate rock lithology identification model in step (3); based on the fact that the carbonate rock is mainly composed of calcite and dolomite, the identification model is used to predict the content of calcite and dolomite in each carbonate rock sample; the measured values ​​of the calcite and dolomite contents in the carbonate rock are obtained by X-ray diffraction detection; and according to the predicted values ​​and measured values ​​of the calcite and dolomite contents, a fitting curve and its absolute coefficient and root mean square error are obtained, and the carbonate rock lithology is identified and classified into two or four categories.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] (1) The method of the present invention is based on the hyperspectral data of carbonate rock samples. After preprocessing, multiple feature selection methods are used for feature selection, especially the random forest method is used to select features with high importance. Based on this, a lithology recognition model is established, and the model parameters are determined by machine learning. The content of mineral components is predicted by the lithology recognition model, thereby realizing the recognition and classification of carbonate rock lithology. The accuracy of lithology recognition can reach more than 80%;

[0062] (2) Compared with manual identification, the method of the present invention greatly improves work efficiency and provides a basis for geological detection of oil and gas reservoirs and oil and gas exploration and development, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is a graph of abnormal point detection results of a carbonate rock sample provided in Example 1 of the present invention;

[0064] Figure 2 1 is the original spectrum of carbonate rock samples 1 and 2 provided in Example 1 of the present invention and the spectrum after decontinuum processing;

[0065] Figure 3 is a fitting curve diagram based on the predicted value and the measured value of the calcite content in the carbonate rock sample provided in Example 1 of the present invention;

[0066] Figure 4It is a fitting curve diagram of the predicted value and the measured value based on the dolomite content in the carbonate rock sample provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0067] In order to better illustrate the present invention and facilitate understanding of the technical solution of the present invention, the present invention is further described in detail below. However, the following embodiments are only simple examples of the present invention and do not represent or limit the scope of protection of the present invention. The scope of protection of the present invention shall be subject to the claims.

[0068] The specific implementation of the present invention provides a carbonate rock lithology identification method based on machine learning, and the lithology identification method comprises the following steps:

[0069] (1) Using ASD spectrometer to obtain hyperspectral data of carbonate rock samples and preprocessing the hyperspectral data;

[0070] (2) Using a hybrid feature selection method, first select the mineral characteristic bands in the hyperspectral data, then use the random forest feature selection method to sort the features according to their importance and select the features with high importance;

[0071] (3) constructing a carbonate rock lithology identification model based on the features selected in step (2), wherein model parameters are determined based on a machine learning algorithm;

[0072] (4) Predicting the mineral content in the carbonate rock according to the carbonate rock lithology identification model in step (3), thereby identifying the lithology of the carbonate rock.

[0073] The following are typical but non-limiting embodiments of the present invention:

[0074] Embodiment 1:

[0075] This embodiment provides a carbonate rock lithology identification method based on machine learning, and the lithology identification method includes the following steps:

[0076] (1) ASD spectrometer is used to obtain hyperspectral data of carbonate rock samples. The ASD spectrometer is a portable ground feature spectrometer. The hyperspectral data is remote sensing imaging spectral data with a spectral resolution of nanometers, which is a reflectance spectrum in the band range of 350 to 2500 nm. The acquisition of the hyperspectral data needs to be calibrated by whiteboard reflectance. The carbonate rock samples include 206 samples selected from eastern and western Sichuan;

[0077] The hyperspectral data is preprocessed, and the preprocessing includes outlier detection and removal and continuum removal. The outlier detection and removal is to remove the sample spectra with abnormal hyperspectral data from the hyperspectral data of many samples. The detection of the outliers adopts the upper and lower cutoff point method in statistics. The selection of the upper and lower cutoff points adopts the quartile method for calculation, wherein the quartiles include the first quartile Q1, the median Q2 and the third quartile Q3, the value of the upper cutoff point is Q3+1.5×(Q3-Q1), and the value of the lower cutoff point is Q1-1.5×(Q3-Q1); the specific calculation process is as follows:

[0078] Step 1: Calculate the upper and lower cutoff values ​​of each band in the 2151 bands within the range of 350-2500nm of the hyperspectral data of 206 samples;

[0079] Step 2: Set the initial value i=1, num=0;

[0080] Step 3: Compare the reflectivity of the sample's i-th band with the upper and lower cutoff points of the band obtained in the first step. If the reflectivity is above the upper threshold or below the lower threshold, it is counted as an abnormal point, num=num+1;

[0081] Step 4: i=i+1, repeat the method in step 3, perform outlier analysis on the reflectance of all bands of the sample in turn, and finally obtain num as the total number of outliers;

[0082] Step 5: If num>2151*0.75, the sample is considered an abnormal sample;

[0083] Step 6: Repeat steps 2 to 5 to complete outlier detection for all samples.

[0084] After calculation, there are 3 abnormal samples detected in 206 samples, and the number of samples to be analyzed is 203. The abnormal point detection results of the carbonate rock samples are shown in the figure Figure 1 As shown;

[0085] The decontinuum processing includes the generation of a continuum and the normalization processing of the continuum elimination method. The Clark shell coefficient method is used to generate a continuum for the spectral curve. The normalization processing of the continuum elimination method includes: using the actual reflectivity of the sample to remove the reflectivity of the corresponding band on the continuum to obtain the value normalized by the continuum elimination method; the original spectra of carbonate rock samples 1 and 2 and the spectra after decontinuum processing are as follows: Figure 2 As shown;

[0086] (2) A hybrid feature selection method is used. First, a filtering method is used to select the feature interval 1800-2400nm in the hyperspectral data, and it is divided into four feature intervals 1800-1920nm, 1920-2020nm, 2060-2250nm and 2250-2400nm. The above feature intervals correspond to mineral characteristic bands. Then, a random forest feature selection method is used to sort the feature importance of the feature bands in the feature intervals, and the feature bands with high importance are selected. The feature bands are sorted by importance using a random forest regression model. Among the parameters of the RandomForestRegressor model, n_estimators is 500, max_depth is 5, and min_samples_leaf is 5. The feature_importance parameter is called to obtain the importance ranking of 560 feature bands, and the feature bands ranked in the top 50 are selected as features with high importance.

[0087] (3) constructing a carbonate rock lithology identification model based on the characteristic bands of high importance selected in step (2), wherein the model parameters are determined based on a machine learning algorithm. During the machine learning, 162 of the 203 samples are used as training samples, and the remaining 41 samples are used as test samples. The machine learning adopts an SVM algorithm, and specifically selects a support vector regression SVR method in the sklearn library. The support vector regression adopts a grid search and cross-validation method to determine a hyperparameter C of the algorithm model. According to the range of the SVR hyperparameter of 0.001 to 200, the step value is adjusted, and the SVR learner is trained using the adjusted parameters. The parameter with the highest accuracy in the validation set is selected as the hyperparameter of the model, wherein the cross-validation parameter is 5, and the SVR hyperparameter C is finally determined to be 150;

[0088] (4) Predicting the mineral content in the carbonate rock according to the carbonate rock lithology identification model in step (3), based on the fact that the carbonate rock is mainly composed of calcite and dolomite, the identification model is used to predict the content of calcite and dolomite in each carbonate rock sample, the measured values ​​of the calcite and dolomite content in the carbonate rock are obtained by X-ray diffraction detection, and the fitting curves are obtained according to the predicted values ​​and measured values ​​of the calcite and dolomite contents, respectively, as shown in FIG. Figure 3 and Figure 4 As shown, the lithology of carbonate rocks is identified according to the relationship between the calcite and dolomite contents. For details, see the carbonate rock lithology classification description table in Table 1. It is divided into two categories of dolomite and limestone and four categories of dolomite, lime-bearing dolomite, dolomite-bearing limestone and limestone. The two-category confusion matrix and four-category confusion matrix of carbonate rock lithology are shown in Tables 1 and 2, respectively.

[0089] Table 2 Confusion matrix of carbonate rock lithology binary classification in Example 1

[0090]

[0091] Table 3 Confusion matrix of four classifications of carbonate rock lithology in Example 1

[0092]

[0093] In this embodiment, Figure 3 It can be seen that the absolute coefficient R in the fitting curve between the predicted value and the measured value of calcite content in carbonate rocks is 2 is 0.889, and the root mean square error RMSE is 0.144; Figure 4 It can be seen that the absolute coefficient R in the fitting curve between the predicted value and the measured value of dolomite content in carbonate rocks is 2 is 0.917, and the root mean square error RMSE is 0.122;

[0094] From the data in Table 2, it can be seen that the calculation process of the overall accuracy is to add the hyperparameters on the diagonal of the confusion matrix and divide them by the total number of samples, that is, (22+15) / 41, and the accuracy of the carbonate rock lithology binary classification is 0.9024; from the data in Table 3, it can be seen that the calculation process of the overall accuracy is (18+0+1+15) / 41, and the accuracy of the carbonate rock lithology quadrature classification is 0.8293.

[0095] Embodiment 2:

[0096] This embodiment provides a carbonate rock lithology identification method based on machine learning, and the lithology identification method includes the following steps:

[0097] (1) ASD spectrometer is used to obtain hyperspectral data of carbonate rock samples. The ASD spectrometer is a portable ground feature spectrometer. The hyperspectral data is remote sensing imaging spectral data with a spectral resolution of nanometers, which is a reflectance spectrum in the band range of 350 to 2500 nm. The acquisition of the hyperspectral data needs to be calibrated by whiteboard reflectance. The carbonate rock samples include 274 samples selected from eastern and western Sichuan;

[0098] Preprocessing the hyperspectral data includes outlier detection and removal and continuum removal. The outlier detection and removal is to remove the sample spectra with abnormal hyperspectral data from the hyperspectral data of many samples. The detection of the outliers adopts the upper and lower cutoff point method in statistics. The selection of the upper and lower cutoff points adopts the quartile method for calculation, wherein the quartiles include the first quartile Q1, the median Q2 and the third quartile Q3. The value of the upper cutoff point is Q3+1.5×(Q3-Q1), and the value of the lower cutoff point is Q1-1.5×(Q3-Q1). After calculation, there are 2 samples with abnormal detection among the 274 samples, and the number of samples to be analyzed is 272 in the end.

[0099] The decontinuum processing includes the generation of a continuum and the normalization processing of the continuum elimination method, and the Clark shell coefficient method is used to generate a continuum for the spectral curve. The normalization processing of the continuum elimination method includes: using the actual reflectivity of the sample to remove the reflectivity of the corresponding band on the continuum to obtain the normalized value of the continuum elimination method;

[0100] (2) A hybrid feature selection method is used. First, a filtering method is used to select the feature interval 1800-2400nm in the hyperspectral data, and four feature intervals 1800-1920nm, 1920-2020nm, 2060-2250nm and 2250-2400nm are extracted. The feature interval corresponds to the mineral characteristic band. Then, a random forest feature selection method is used to sort the feature importance of the feature bands in the feature interval, and the feature bands with high importance are selected. The feature bands are sorted by importance using a random forest regression model. Among the parameters of the RandomForestRegressor model, n_estimators is 500, max_depth is 5, and min_samples_leaf= is 5. The feature_importance parameter is called to obtain the importance ranking of 560 feature bands, and the feature bands ranked in the top 50 are selected as features with high importance.

[0101] (3) constructing a carbonate rock lithology identification model based on the characteristic bands with high importance selected in step (2), wherein the model parameters are determined based on a machine learning algorithm, wherein the machine learning adopts an SVM algorithm, and specifically selects a support vector regression SVR method in the sklearn library, and the specific algorithm and calculation process adopted are the same as those in Example 1, and the obtained hyperparameters are also the same, and the above 272 samples to be analyzed can all be used as test samples;

[0102] (4) Predicting the mineral content in the carbonate rock according to the carbonate rock lithology identification model in step (3), based on the fact that the carbonate rock is mainly composed of calcite and dolomite, the identification model is used to predict the content of calcite and dolomite in each carbonate rock sample, the measured values ​​of the calcite and dolomite content in the carbonate rock are obtained by XRD detection, and according to the relationship between the calcite and dolomite contents, the carbonate rock lithology classification description table in Table 1 can be specifically seen, and the lithology of the carbonate rock is identified, and the carbonate rock is classified according to the two-classification of dolomite and limestone and the four-classification of dolomite, lime-bearing dolomite, dolomite-bearing limestone and limestone. That is, the carbonate rock lithology two-classification confusion matrix and four-classification confusion matrix are shown in Table 4 and Table 5, respectively.

[0103] Table 4 Confusion matrix of carbonate rock lithology binary classification in Example 2

[0104]

[0105] Table 5 Confusion matrix of four classifications of carbonate rock lithology in Example 2

[0106]

[0107] It can be seen from the above embodiments that the method of the present invention is based on the hyperspectral data of carbonate rock samples, and after preprocessing, a hybrid feature selection method is used for feature selection, especially a random forest method is used to select features with high importance, and a lithology recognition model is established on this basis, and the model parameters are determined by machine learning. The content of mineral components is predicted by the lithology recognition model, thereby realizing the recognition and classification of carbonate rock lithology, and the accuracy of lithology recognition can reach more than 80%. Compared with the manual recognition method, the method greatly improves the work efficiency, provides a basis for geological detection of oil and gas reservoirs and oil and gas exploration and development, and has broad application prospects.

[0108] The applicant declares that the present invention illustrates the detailed method of the present invention through the above-mentioned embodiments, but the present invention is not limited to the above-mentioned detailed method, that is, it does not mean that the present invention must rely on the above-mentioned detailed method to be implemented. Those skilled in the art should understand that any improvement of the present invention, equivalent replacement of the method of the present invention, addition of auxiliary methods, selection of specific methods, etc., all fall within the protection scope and disclosure scope of the present invention.

Claims

1. A carbonate rock lithology identification method based on machine learning, characterized in that: The lithology identification method comprises the following steps: (1) Using ASD spectrometer to obtain hyperspectral data of carbonate rock samples and preprocessing the hyperspectral data; (2) Using a hybrid feature selection method, first select the mineral characteristic bands in the hyperspectral data, then use the random forest feature selection method to sort the features according to their importance and select the features with high importance; (3) constructing a carbonate rock lithology identification model based on the features selected in step (2), wherein model parameters are determined based on a machine learning algorithm; (4) Predicting the mineral content in the carbonate rock according to the carbonate rock lithology identification model in step (3), thereby identifying the lithology of the carbonate rock.

2. The lithology identification method according to claim 1, characterized in that: The ASD spectrometer in step (1) is a portable ground object spectrometer; Preferably, the hyperspectral data in step (1) is remote sensing imaging spectral data with a spectral resolution of nanometers; Preferably, the hyperspectral data in step (1) is a reflectance spectrum in the wavelength range of 350 to 2500 nm; Preferably, the acquisition of the hyperspectral data in step (1) needs to be corrected by the reflectivity of a white board.

3. The lithology identification method according to claim 1 or 2, characterized in that: The preprocessing in step (1) includes outlier detection and elimination and decontinuum processing; Preferably, the abnormal point detection and elimination is to eliminate the sample spectra with abnormal hyperspectral data from the hyperspectral data of a large number of samples; Preferably, the detection of the abnormal points adopts the upper and lower cutoff point method in statistics; Preferably, the selection of the upper and lower cutoff points is calculated using the quartile method, wherein the quartiles include the first quartile, the median and the third quartile.

4. The lithology identification method according to claim 3, characterized in that: The decontinuum processing in the preprocessing includes the generation of a continuum and a normalization processing using a continuum elimination method; Preferably, the Clark shell coefficient method is used to generate a continuum of the spectral curve; Preferably, the normalization processing by the continuum elimination method comprises: using the actual reflectivity of the sample to remove the reflectivity of the corresponding band on the continuum to obtain a normalized value by the continuum elimination method.

5. The lithology identification method according to any one of claims 1 to 4, characterized in that: The hybrid feature selection method in step (2) includes a combination of at least two of a filtering method, an embedded method or an encapsulation method, preferably a combination of a filtering method and an encapsulation method; Preferably, the filtering method is to select the characteristic interval of 1800-2400nm in the hyperspectral data and divide it into four characteristic intervals of 1800-1920nm, 1920-2020nm, 2060-2250nm and 2250-2400nm; Preferably, the characteristic intervals of 1800-1920nm, 1920-2020nm, 2060-2250nm and 2250-2400nm correspond to mineral characteristic bands.

6. The lithology identification method according to claim 5, characterized in that: The encapsulated method is a feature selection method based on random forests, which sorts the feature bands in the feature interval by importance and selects the feature bands with high importance; Preferably, the characteristic bands are ranked by importance using a random forest regression model.

7. The lithology identification method according to any one of claims 1 to 6, characterized in that: The features of the recognition model constructed in step (3) are the feature bands with high importance selected in step (2); Preferably, during the machine learning in step (3), 70-90% of the samples are used as training samples and the remaining samples are used as test samples.

8. The lithology identification method according to any one of claims 1 to 7, characterized in that: The machine learning algorithm in step (3) adopts a support vector machine (SVM) algorithm, preferably a support vector regression (SVR) method; Preferably, the support vector regression uses a grid search and cross-validation method to determine a hyperparameter C of the algorithm model; Preferably, the SVR hyperparameters are adjusted in step size according to their range, and the adjusted parameters are used for training, and the parameters with the highest accuracy in the validation set are selected as the hyperparameters of the model.

9. The lithology identification method according to any one of claims 1 to 8, characterized in that: Based on the fact that the carbonate rock is mainly composed of calcite and dolomite, the identification model is used to predict the content of calcite and dolomite in each carbonate rock sample; Preferably, the measured values ​​of the calcite and dolomite contents in the carbonate rock are obtained by X-ray diffraction detection; Preferably, a fitting curve and its absolute coefficient and root mean square error are obtained based on the predicted values ​​and measured values ​​of the calcite and dolomite contents in the carbonate rock; Preferably, the lithology classification of the carbonate rock sample is determined based on the predicted values ​​of the calcite and dolomite contents in the carbonate rock sample; Preferably, the lithology classification of the carbonate rock is divided into two categories and four categories, the two categories are dolomite and limestone, and the four categories are dolomite, lime-bearing dolomite, dolomite-bearing limestone and limestone.

10. The lithology identification method according to any one of claims 1 to 9, characterized in that: The lithology identification method comprises the following steps: (1) ASD spectrometer is used to obtain hyperspectral data of carbonate rock samples. The ASD spectrometer is a portable ground feature spectrometer. The hyperspectral data is remote sensing imaging spectral data with a spectral resolution of nanometers, which is a reflectance spectrum in the band range of 350 to 2500 nm. The acquisition of the hyperspectral data needs to be corrected by the reflectance of a whiteboard; and the hyperspectral data is preprocessed, and the preprocessing includes outlier detection and removal and continuum removal. The outlier detection and removal is to remove the sample spectra with abnormal hyperspectral data from the hyperspectral data of many samples. The detection of the outliers adopts the upper and lower cutoff point method in statistics, and the selection of the upper and lower cutoff points is calculated by the quartile method; the continuum removal includes the generation of a continuum and the normalization of the continuum elimination method. The Clark shell coefficient method is used to generate a continuum for the spectral curve. The continuum elimination normalization includes: using the actual reflectance of the sample to remove the reflectance of the corresponding band on the continuum to obtain the normalized value of the continuum elimination method; (2) A hybrid feature selection method is used. A filtering method is first used to select a feature interval of 1800-2400 nm in the hyperspectral data, and the interval is divided into four feature intervals: 1800-1920 nm, 1920-2020 nm, 2060-2250 nm, and 2250-2400 nm. The four feature intervals correspond to mineral feature bands. A random forest feature selection method is then used to sort the feature bands in the feature intervals by feature importance, and feature bands with high importance are selected. The feature bands are sorted by importance using a random forest regression model. (3) constructing a carbonate rock lithology identification model based on the characteristic bands of high importance selected in step (2), wherein the model parameters are determined based on a machine learning algorithm. During the machine learning, 70 to 90% of the samples are used as training samples, and the remaining samples are used as test samples. The machine learning algorithm adopts a support vector regression (SVR) method, and the support vector regression uses a grid search and cross-validation method to determine the optimal hyperparameter C of the algorithm model. According to the range of the SVR hyperparameter, the step size is adjusted, and the adjusted parameter is used for training, and the parameter with the highest accuracy in the validation set is selected as the hyperparameter of the model; (4) Predicting the mineral content in the carbonate rock according to the carbonate rock lithology identification model in step (3); based on the fact that the carbonate rock is mainly composed of calcite and dolomite, the identification model is used to predict the content of calcite and dolomite in each carbonate rock sample; the measured values ​​of the calcite and dolomite contents in the carbonate rock are obtained by X-ray diffraction detection; and according to the predicted values ​​and measured values ​​of the calcite and dolomite contents, a fitting curve and its absolute coefficient and root mean square error are obtained, and the carbonate rock lithology is identified and classified into two or four categories.

Images