Lithology automatic classification method and system based on cooperation of hyperspectral image and machine learning

Abstract

The invention discloses an automatic lithology classification method and system based on cooperation of a hyperspectral image and machine learning. The method comprises the following steps: collecting training sample pattern spots, test samples and rock samples in the field; measuring a typical spectral curve of the rock sample for researching spectral curve characteristics corresponding to each type of lithology; performing two-stage dimension reduction processing on the hyperspectral image by using the characteristics of the typical spectral curve; a classification model is constructed by utilizing machine learning, the hyperspectral image after dimension reduction is classified by utilizing the classification model, and lithology classification result pattern spots are generated; and verifying and evaluating the classification precision of the lithology classification result pattern spots. According to the method, physical priori knowledge of a rock actually measured spectrum and an advanced data driving algorithm are combined, so that the limitation of an existing dimension reduction method can be effectively solved, the classification precision and reliability of a machine learning model in a complex bare rock area can be improved, and an efficient and accurate new way is provided for lithology classification of a hyperspectral remote sensing image.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing geology and machine learning, and in particular to a lithology automatic classification method based on hyperspectral image and machine learning. BACKGROUND

[0002] Hyperspectral remote sensing technology provides an important means for fine lithology identification. Although it can obtain fine spectral characteristics of ground objects and provide a rich data basis for lithology classification, it also faces the problem of "dimension disaster" caused by hundreds of bands, that is, the data volume is too large and there is a lot of redundant information, which leads to low classification efficiency. In order to deal with the "dimension disaster", traditional dimension reduction methods are mainly divided into two categories: feature transformation-based methods (such as PCA, LDA) and band selection-based methods (such as correlation coefficient method, information entropy method). The former may destroy the original spectral physical meaning, and the latter such as SVM-RFE has high computational complexity. In recent years, the Pearson correlation coefficient method combined with spectral physical characteristics has begun to attract attention in preserving the diagnostic absorption characteristics of rocks, providing a new idea for lithology classification. However, in the actual application of bare rock area, the dimension reduction still has the problems of insufficient sensitivity to weak spectral differences and being easily disturbed by terrain shadows, and it is urgent to develop an efficient dimension reduction strategy.

[0003] At the same time, the traditional lithology interpretation method mainly relies on manual visual interpretation and multispectral data, and the classification accuracy is limited by the lack of wave bands and the experience of interpreters. With the rise of machine learning technology, the existing hyperspectral lithology classification method also has limitations: deep learning models, such as convolutional neural networks (CNN), although perform well in complex scene classification, but due to the difficulty in obtaining samples and poor model interpretability, their application in bare rock area is limited to a certain extent; in addition, most of the existing methods directly use full-band data or simple dimension reduction results as input, without fully combining the physical prior knowledge of rock measured spectrum, which leads to misclassification phenomenon in the lithology spectral overlapping area.

[0004] In summary, although hyperspectral remote sensing technology has great potential in lithology classification, it also faces many challenges. The existing dimension reduction method and machine learning model cannot meet the needs of high-precision lithology classification. Therefore, how to combine physical mechanism and data-driven to build a more robust and accurate classification framework has become an important research direction. SUMMARY

[0005] The embodiment of the present application provides a lithology automatic classification method and system based on hyperspectral image and machine learning to solve the limitations of the existing dimension reduction method and the technical problem of how to realize efficient dimension reduction and significantly improve the classification accuracy and efficiency of bare rock area.

[0006] A lithology automatic classification method based on hyperspectral image and machine learning, comprising:

[0007] training sample patches, test samples and rock samples in the field;

[0008] measuring typical spectral curves of the rock samples for studying spectral curve features corresponding to each type of lithology;

[0009] performing two-stage dimension reduction processing on the hyperspectral image by using the features of the typical spectral curves;

[0010] constructing a classification model by using machine learning, classifying the dimension-reduced hyperspectral image by using the classification model, and generating a lithology classification result patch;

[0011] verifying and evaluating the classification accuracy of the lithology classification result patch.

[0012] A lithology automatic classification system based on cooperation of hyperspectral image and machine learning, comprising:

[0013] a collection module for collecting training sample patches, test samples and rock samples in the field;

[0014] a spectral measurement module for measuring typical spectral curves of the rock samples for studying spectral curve features corresponding to each type of lithology;

[0015] a dimension reduction processing module for performing two-stage dimension reduction processing on the hyperspectral image by using the features of the typical spectral curves;

[0016] a classification modeling module for constructing a classification model by using machine learning, classifying the dimension-reduced hyperspectral image by using the classification model, and generating a lithology classification result patch;

[0017] an accuracy verification module for verifying and evaluating the classification accuracy of the lithology classification result patch.

[0018] The present application has the following beneficial effects:

[0019] (1) The application adopts a multi-dimensional data dimension reduction strategy, which is different from the single dimension reduction technology or pure statistical band selection commonly used in traditional methods, and the traditional method often easily leads to the loss of effective spectral features or the residual of redundant information. The two-stage dimension reduction system of "field knowledge guidance + statistical optimization" proposed in the application skillfully combines geological knowledge and statistical methods. In the first stage, by comparing and observing the rock spectrum curve, the spectrum band with obvious difference and its surrounding band are selected as the main information carrier band, and the key band is selected according to the difference characteristics of the measured spectrum curve, so that the sensitive spectrum band which is crucial to geological analysis is accurately retained; in the second stage, the Pearson correlation coefficient method with dynamic threshold is used, the correlation between each band and other bands is calculated by iterative screening, the bands with low correlation are selected for use, the nonlinear redundancy is removed, and the most representative bands are retained, so that the effective information is retained to the greatest extent while the dimension is reduced, and the perfect balance between dimension reduction and information retention is realized, thereby providing a high-quality data basis for subsequent lithology classification.

[0020] (2) The application also makes significant progress in improving the correlation between spectrum and ground objects. The traditional technology often directly depends on the laboratory standard spectrum library, but ignores the offset influence of surface weathering on spectrum. In the sampling and spectrum measurement process, more attention is paid to the characteristics of the rock weathering surface, thereby significantly improving the correlation between the ground object spectrum and the image spectrum, enabling the classification model to more accurately identify and classify different lithology, thereby improving the reliability and accuracy of the classification result.

[0021] (3) The application implements full-process quality control in the two key links of training sample collection and classification result verification, ensuring the accuracy and reliability of the classification result. In the collection of training samples, accurate samples drawn in the field are used instead of simply drawing a few areas from the existing geological map, avoiding the problem of inaccurate sample data caused by the mapping error and geological mapping unit merging of the geological map, and ensuring the accuracy and representativeness of the training samples. At the same time, in the verification of the classification result, the small-scale geological mapping result in the field is used as the evaluation standard instead of using the test set samples reserved in proportion in the training samples for evaluation, avoiding the evaluation accuracy deviation caused by the high similarity between the training set and the test set data, and making the evaluation result more real and valuable. Through the above method, the performance of the classification model can be accurately evaluated, and the reliability and practicality of the classification result can be ensured. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed to be used in the description of the embodiments of the application will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creating any inventive labor.

[0023] Figure 1 This is a flowchart of an automatic lithology classification method that combines hyperspectral imagery and machine learning in one embodiment of the present invention;

[0024] Figure 2 This is a schematic diagram of a lithological automatic classification system that combines hyperspectral imagery with machine learning in one embodiment of the present invention.

[0025] Figure 3 This is another schematic diagram of an automatic lithology classification system that combines hyperspectral imagery with machine learning in one embodiment of the present invention;

[0026] Figure 4 This is a schematic diagram of a computer device for automatic lithological classification that combines hyperspectral imagery with machine learning, according to an embodiment of the present invention.

[0027] Figure 5 This is a distribution map of lithological samples according to an embodiment of the present invention;

[0028] Figure 6 This is a small-scale geological map according to an embodiment of the present invention;

[0029] Figure 7 This is a schematic diagram of the mean removal of a rock spectral curve according to an embodiment of the present invention;

[0030] Figure 8 This is a schematic diagram of a remote sensing image after mask extraction according to an embodiment of the present invention;

[0031] Figure 9 This is a schematic diagram of a spectral curve comparison and selection of a band according to an embodiment of the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] The lithological automatic classification method based on hyperspectral imagery and machine learning provided in this invention can be applied to, for example... Figure 1 The application environment is shown. Specifically, this lithological automatic classification method combining hyperspectral imagery and machine learning is applied in a lithological automatic classification system combining hyperspectral imagery and machine learning, which includes, for example, […]. Figure 3The illustrated client and server communicate through a network, and can improve the classification accuracy and efficiency of bare rock areas while efficiently reducing the dimensionality, and can overcome the limitations of existing dimensionality reduction methods. The client, also known as the user end, is a program that provides local services for the client corresponding to the server. The client can be on various personal computers. The server can be implemented by a stand-alone server or a server cluster composed of multiple servers.

[0034] In an embodiment, as Figure 1 illustrated, a hyperspectral image and machine learning collaborative lithology automatic classification method is provided. Since the ZY1F hyperspectral remote sensing image has continuous subdivided spectrum (166 bands), its spectral resolution is much higher than that of multispectral remote sensing images (such as Landsat8 images). The improvement of spectral resolution makes it possible to provide more detailed spectral information of ground object reflection curve, which provides a rich spectral information basis for fine geological interpretation tasks such as lithology identification. However, the continuous subdivision characteristics of spectral data also bring double challenges: on the one hand, there is a high degree of redundant information between adjacent bands, which leads to an increase in data dimension, directly increasing the computational load of subsequent classification processing and reducing work efficiency; on the other hand, the noise and non-diagnostic information mixed in the redundant bands may interfere with the extraction of key geological features, reducing the classification efficiency. The hyperspectral image and machine learning collaborative lithology automatic classification method aims to achieve two goals: first, in the process of reducing the dimensionality of hyperspectral image data, the effective information is preserved to the greatest extent, so that when performing the classification task, the computational load can be reduced, and the integrity of the effective information can be ensured to the greatest extent; second, a lightweight machine learning method is introduced to replace the traditional remote sensing geological work process that relies on manual visual interpretation, effectively avoiding errors and information interpretation limitations caused by manual operation. Take the server in the method applied in Figure 3 as an example for illustration, including the following steps:

[0035] S1, field collection of training sample polygons, test samples and rock samples.

[0036] In an embodiment, the step S1 of field collection of training sample polygons, test samples and rock samples further includes the following substeps:

[0037] S101, analyze the basic geological data of the study area to determine the lithology types and quantities of the study area. Provide directional guidance for subsequent sampling work.

[0038] S102, carry out field reconnaissance, determine the typical distribution area of each type of lithology, and draw the training sample polygons. These training samples do not need to cover the entire rock mass boundary, but need to ensure that the training samples are evenly distributed in the study area. If conditions permit, they should be dispersed as much as possible to enhance the representativeness of the samples.

[0039] S103, sample collection is performed on each type of rock, and a rock sample with a natural weathered surface is collected. Since the satellite image reflects the characteristics of the weathered rock on the ground, the block picking method is used when collecting the physical sample of each type of rock, and the rock block with the natural weathered surface is collected to improve the similarity between the measured spectrum and the image spectrum. Multiple physical samples of each lithology should be collected, and the sampling position should be dispersed to avoid spatial clustering effect and be uniformly distributed in the study area to ensure the diversity and representativeness of the samples.

[0040] S104, a typical area is selected and geological lithology mapping is performed as a test sample, and the accurate distribution boundary of the rock mass is outlined. According to the outcrop condition and traffic condition of the rock mass in the study area, blocks with good bedrock exposure, typical lithology and covering all lithologies in the area are selected. These blocks do not have to be adjacent, but each lithology in the area must be included. Then, small-scale geological mapping is performed on the small-scale blocks, and the accurate distribution boundary of the rock mass is accurately outlined. This part of data will be used as a verification benchmark for the subsequent classification result evaluation to ensure the accuracy and reliability of the classification result.

[0041] S2, the typical spectral curve of the rock sample is measured, which is used to study the spectral curve characteristics corresponding to each type of lithology. Understandably, the typical spectral curve of the rock sample is measured, which can realize the standardization measurement of the spectral characteristics of the rock.

[0042] In an embodiment, the step S2 of measuring the typical spectral curve of the rock sample, which is used to study the spectral curve characteristics corresponding to each type of lithology, further includes the following sub-steps:

[0043] S201, multiple multi-angle spectral curve measurements are performed on each rock sample collected, and the spectral curve of each rock sample and its corresponding lithology type are recorded during the measurement process. The spectral range of the used spectrometer should be greater than the spectral range of the hyperspectral remote sensing image to be classified, so as to ensure that the measured spectral curve can cover all the band information of the image. Each sample can be measured multiple times from multiple angles. The spectral curve of each sample and its corresponding lithology type are recorded during the measurement process, and the random error is reduced through multiple measurements to obtain more comprehensive and accurate spectral information.

[0044] S202, the average value of the spectral curves measured for all samples of each type of rock is taken as the typical spectral curve of the rock in the area. The typical spectral curve is used to study the spectral curve characteristics corresponding to each type of lithology. This step helps to reduce measurement error and improve the representativeness of the spectral curve, providing a more accurate reference for subsequent image processing and classification.

[0045] S3, the characteristics of the typical spectral curve are used for two-stage dimension reduction processing of the hyperspectral image.

[0046] In an embodiment, the step S3 utilizes the characteristics of the typical spectral curve to perform two-stage dimension reduction processing on the hyperspectral image, and further includes the following sub-steps:

[0047] S301, pre-processing the original image, the pre-processing including radiation calibration, atmospheric correction, geometric correction and other standardized preprocessing operations to ensure data quality. Understandably, the purpose of this operation is to eliminate various interference factors that the image has suffered during acquisition, and to improve the quality and availability of the image. Radiation calibration can convert the digitized value of the image into a physical quantity, atmospheric correction can eliminate the influence of the atmosphere on the image, and geometric correction can correct the geometric distortion of the image to make it consistent with the actual geographical coordinates.

[0048] S302, combined with the terrain and high-resolution remote sensing image, the Quaternary coverage area of the study area is circled out, and the Quaternary coverage area image is removed by mask cropping to exclude the interference of non-rock targets. Understandably, since the gravel and soil components of the Quaternary coverage area are often very complex, and are often formed by multiple sources, in order to exclude the influence of the Quaternary on the classification work, the image needs to be cropped using the circled Quaternary coverage area to only keep the image of the non-Quaternary coverage area. This step helps to reduce interference information and improve the accuracy of classification.

[0049] S303, entering the first stage of dimension reduction, comparing the typical spectral curve of each type of rock sample, and manually selecting the spectral frequency band and its adjacent wave band sensitive to lithology classification and identification based on the difference characteristics of the measured spectral curve. Understandably, by comparing the typical spectral curve of each type of rock collected in step S2, the absorption peak, absorption valley and wave band with large reflectivity difference of different types of rock on the spectral curve of each type of rock are selected, and then 3-5 wave bands are taken forward and backward from each characteristic wave band as the main information carrier wave band. Based on the difference characteristics of the measured spectral curve of the rock, this step manually preliminarily selects the key wave band important to lithology classification from the original image, providing a basis for subsequent dimension reduction processing.

[0050] S304, entering the second stage of dimension reduction, using the Pearson correlation coefficient (Pearson) method for dynamic threshold dimension reduction based on the manual preliminary selection.

[0051] Further, in an embodiment, the step S304 further includes the following sub-steps:

[0052] S3041, by comparing the correlation coefficients between the wave band data, the correlation of each wave band with other wave bands is calculated, and the wave band with smaller correlation is selected for use. Understandably, the Pearson correlation coefficient of each wave band relative to other wave bands is calculated one by one, the absolute value of the calculated correlation coefficient is taken and then added, and then sorted to select the wave band with the smallest correlation coefficient added with all other wave bands to join the selected wave band list.

[0053] S3042, cycle selection of wave bands, in each cycle, calculate the average correlation between the remaining wave bands and the selected wave bands, and select the wave band with the lowest correlation to join the selected wave band list. Understandably, the most representative wave band is gradually screened out to achieve effective dimension reduction of hyperspectral image and ensure that the information redundancy between the selected wave bands is minimized.

[0054] S3043, until the cycle selection of wave bands reaches the maximum preset reserved wave band number or the correlation of all remaining wave bands is higher than a preset threshold of a certain correlation, then stop selecting wave bands, and synthesize the reduced wave bands into a new GeoTIFF file. Understandably, this step removes redundant information and retains effective information through iterative screening, effectively solving the "dimension disaster" problem of hyperspectral image and improving the processability and classification efficiency of the image.

[0055] S4, constructing a classification model using machine learning, classifying the reduced hyperspectral image using the classification model, and generating a lithology classification result polygon.

[0056] In an embodiment, the step S4 of constructing a classification model using machine learning, classifying the reduced hyperspectral image using the classification model, and generating a lithology classification result polygon further includes the following sub-steps:

[0057] S401, using the field-drawn training sample polygons to extract corresponding remote sensing image spectral values to construct classification samples. Understandably, the typical lithology area (training sample polygons) drawn in step S1 is converted into a point file for extracting the pixel values of the remote sensing image. Since the spatial resolution of ZY1F hyperspectral image is 30 meters, the point distance of the point file generated from the area should be greater than 30 meters to ensure that each point extracts the value of a different pixel, avoiding repeated sample points. The purpose of this step is to obtain spectral samples with clear lithology labels to provide a data basis (training set) for the training of the machine learning model to ensure the spatial consistency of the sample and the model input features.

[0058] S402, selecting random forest (RF), support vector machine (SVM), or XGboost algorithm to model the classification samples, and adjusting and optimizing the hyperparameters of the model through grid search (Grid Search) combined with cross-validation method (Cross-Validation). Understandably, grid search (Grid Search) can traverse all specified parameter combinations to find the optimal parameter settings; cross-validation method (Cross-Validation) can effectively evaluate the performance of the model to prevent overfitting. Through the combination of these two methods, the model can achieve the best performance during the training process to balance the model complexity and generalization ability.

[0059] S403, training a classification model using the optimal hyperparameter combination, and using the classification model to perform end-to-end inference on the dimensionality-reduced hyperspectral image to generate a lithology classification result map. Understandably, this step is to apply the trained model to the hyperspectral image of the entire study area to achieve automatic classification of lithology. By using a lightweight machine learning method, instead of the traditional manual visual interpretation process of remote sensing geological work, errors in visual interpretation and limitations in information interpretation can be effectively avoided, and the efficiency and accuracy of classification can be improved.

[0060] S5, verifying and evaluating the classification accuracy of the lithology classification result map.

[0061] In an embodiment, the step S5 of verifying and evaluating the classification accuracy of the lithology classification result map further comprises the following sub-steps:

[0062] S501, comparing the result map of the geological lithology mapping and the lithology classification result map. Understandably, this step compares the accurate lithology distribution information obtained through field investigation with the model classification result to evaluate the accuracy of the classification result.

[0063] S502, calculating the overall classification accuracy and Kappa coefficient through the confusion matrix, for comparing and evaluating the lithology classification results of different machine learning algorithms, and evaluating the classification errors of different lithology types based on the calculated overall classification accuracy, Kappa coefficient and other quantitative indicators. Understandably, the overall classification accuracy reflects the consistency degree of the classification result and the actual lithology distribution, and the Kappa coefficient considers the random consistency of the classification result, which can more objectively evaluate the reliability of the classification result. Through the calculation of these two evaluation indicators, the performance of different machine learning algorithms in lithology classification can be comprehensively evaluated, providing a basis for selecting the optimal classification algorithm.

[0064] S503, performing spectral feature analysis on misclassified areas to trace the error sources. Understandably, the error sources can be band aliasing, model limitations, etc. By analyzing these factors, the data processing and model construction process can be optimized accordingly, thereby providing a basis for improving the classification method and improving the classification accuracy.

[0065] S504, comparing the performance of different machine learning algorithms, selecting the optimal model for regional promotion, and forming a complete lithology classification technology closed loop. Understandably, this step helps to ensure the accuracy and reliability of the classification result, and at the same time provides solid technical support for subsequent geological research and application.

[0066] It should be understood that the size of the serial number of each step in the above embodiments does not mean the order of execution, and the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0067] In a specific embodiment, as shown in Figures 5 to 9

[0068] Step S1 is specifically: first, comprehensively analyze the regional geological map of the study area, and fully master the lithology types and quantities in the study area. Next, carry out field reconnaissance to determine the distribution position of each type of typical rock. It is not necessary to accurately define the rock mass boundary, but only to circle an accurate typical area on the map or GIS software as a training sample, ensuring that the area is completely of a certain lithology (as shown in Figure 5 If the terrain and traffic conditions permit, try to make the training samples evenly distributed in the study area.

[0069] Then, while circling the sample map of each type of rock, collect typical samples of each type of rock. Since the satellite shot image is the rock on the ground after weathering, when collecting rock samples, the block with weathered surface should be collected to improve the similarity of the measured spectral curve and the spectral curve of the satellite image. Collect more samples of each type of rock and ensure that the sample collection position is dispersed and evenly distributed in the study area.

[0070] Next, according to the outcrop and traffic conditions of the rock mass in the study area, select several typical blocks with good bedrock exposure and lithology. These blocks do not have to be adjacent, but must ensure that every lithology in the area is included. Carry out geological mapping of these small range blocks to accurately delineate the distribution boundary line of each type of rock mass (as shown in Figure 6

[0071] Step S2 is specifically: first, measure the spectral curve of each rock sample collected in step S1, use ispecField-WNIR ground feature spectrometer, its spectral range is 200nm-2500nm, ensure that the spectral range is greater than the spectral range of the hyperspectral remote sensing image to be classified, in order to obtain more comprehensive spectral information.

[0072] Then, measure each rock sample from multiple angles to ensure that the obtained spectral curve can accurately reflect the spectral characteristics of the rock. In the measurement process, record the rock type corresponding to each spectral curve in detail for subsequent analysis.

[0073] Next, after completing the measurement of all samples, process the spectral curves measured for all samples of each type of rock. The specific operation is to calculate the average value of the spectral curves of all samples of each type of rock to eliminate accidental errors of individual measurements, and obtain the typical spectral curve of the rock type in this area (as shown in Figure 7

[0074] Step S3 is specifically: ​​​

[0075] Firstly, the hyperspectral image is processed, including radiation calibration, atmospheric correction and geometric correction, which are carried out according to the traditional data processing procedure to ensure the quality and availability of the image data. Then, the mask cutting work is carried out, which needs to be combined with the terrain and high-resolution remote sensing image to delineate the Quaternary coverage area of the study area. In view of the complex composition and multi-source mixing of the Quaternary coverage area, in order to avoid its interference with the classification work, the image is cut using the delineated Quaternary coverage area, only the image of the non-Quaternary coverage area is retained (as shown in Figure 8 ), so as to improve the accuracy of subsequent analysis.

[0076] Then, the first stage of hyperspectral image dimension reduction is entered, and the spectral curve information of each type of rock collected in step two is compared and observed, and the absorption peaks, absorption valleys and wave bands with large reflectivity difference of different types of rocks (as shown in Figure 9 ) are selected. In order to reduce the difference of artificial subjective selection, 3-5 wave bands are taken in front and back of each characteristic wave band as the main information carrier wave band, which provides more comprehensive reference for subsequent dimension reduction.

[0077] Finally, the second stage of hyperspectral image dimension reduction is carried out, and on the basis of the selection in the first stage, Pearson correlation coefficient method is used for dimension reduction. By calculating the correlation between each wave band and other wave bands (the higher the correlation, the more redundant information), the wave bands with smaller correlation are selected to remove redundant information. Since there is no ready-made hyperspectral image dimension reduction tool based on Pearson correlation coefficient in existing ENVI and Arcgis remote sensing and geographic information software, this step needs to be considered for independent programming to meet the specific dimension reduction demand, and to ensure the scientificity and effectiveness of the dimension reduction process. The principle is as follows: first, calculate the Pearson correlation coefficient of each wave band relative to other wave bands, and the calculation expression is as follows:

[0078]

[0079] Wherein, r represents Pearson correlation coefficient (-1≤r≤1), n represents sample size, x and y represent reflectivity observation values of two wave bands, and Σ represents pixel-level summation operation.

[0080] Take the absolute value of the calculated correlation coefficient and accumulate it to get the sum of the correlation coefficient of each wave band with all other wave bands, and then sort these sums to select the one with the smallest sum of correlation coefficients with all other wave bands to join the selected wave band list.

[0081] Then, the bands are selected in a loop, and in each loop, the average correlation between the remaining bands and the selected bands is calculated, and the band with the lowest correlation is selected to join the selected band list. This process is repeated until one of the following two conditions is met:

[0082] One is that the number of selected bands reaches the maximum number of reserved bands, and the other is that the correlation of all remaining bands is higher than a certain preset threshold. Then, the band selection is stopped, and the reduced bands are combined into a new GeoTIFF file for subsequent analysis and processing.

[0083] Step S4 is specifically: first, build a training sample set, and convert the typical lithology area polygons drawn in step S1 into point files for extracting the pixel values of the remote sensing image. Considering the 30-meter spatial resolution characteristics of ZY1F hyperspectral image, the point spacing is set to be greater than 30 meters to ensure that each sampling point corresponds to an independent pixel, so that each point extracts the value of a different pixel, avoiding sample point repetition.

[0084] Then, machine learning modeling and optimization are implemented, and classic algorithms such as support vector machine (SVM), random forest (RF), or XGboost are selected to build a classification model. During modeling, the grid search (Grid Search) combined with K-fold cross-validation (K-Fold CV) method is used to search the hyperparameter space.

[0085] Next, the best hyperparameter combination in each algorithm is selected to train the model. These models will be used to classify the hyperspectral image. Finally, the trained model is used to classify the entire hyperspectral image to obtain the final classification result.

[0086] Step S5 is specifically: first, the classification result is tested, and the result polygons of the small-scale geological mapping in step S1 are compared with the automatic classification polygons of the remote sensing image. The overall classification accuracy and Kappa coefficient are calculated to quantify the classification effect.

[0087] Then, since existing GIS software usually integrates the functions of calculating classification accuracy and Kappa coefficient in a specific workflow, there is a lack of specialized tools for directly calculating the accuracy and Kappa coefficient between two images, so self-programming is needed to realize this function. When programming is implemented, a programming language such as Python can be used, and functions in third-party libraries such as sklearn can be called to simplify the calculation process. The calculation principle is as follows: when calculating the classification accuracy, the classification image and the measured image value are extracted item by item. For example, let a1_f be the classification image value corresponding to a1 pixel, and a1_s be the measured image value corresponding to a1 pixel. If a1_f is equal to a1_s, that is, the classification is accurate, then the number of pixels with accurate classification is counted. The number of pixels with accurate classification is divided by the total number of pixels, and the result is the overall classification accuracy.

[0088] For the calculation of Kappa coefficient, confusion matrix is used. Confusion matrix is a table used to summarize the difference between the predicted results of the classifier and the actual values. The calculation expression of Kappa coefficient is:

[0089] κ = (po-pe) / (1-pe) (2)

[0090] Where po represents the observed consistency, that is, the proportion of the classifier prediction results and the actual results completely consistent, which can be calculated by dividing the sum of the diagonal elements in the confusion matrix by the total number of samples; pe represents the expected consistency, that is, the consistency that may occur in the case of random classification, which is calculated as the sum of the expected probabilities of each class being correctly predicted.

[0091] The interpretation of Kappa coefficient usually follows the following standards: 0.81-1.00 indicates almost complete consistency; 0.61-0.80 indicates substantial consistency; 0.41-0.60 indicates moderate consistency; 0.21-0.40 indicates mild consistency; 0.0-0.20 indicates almost no consistency. By calculating these two evaluation indicators, the classification results of different machine learning algorithms can be compared and evaluated, so as to select the most suitable classification algorithm.

[0092] In an embodiment, a hyperspectral image and machine learning collaborative lithology automatic classification system is provided, which corresponds to the hyperspectral image and machine learning collaborative lithology automatic classification method in the above embodiment. As shown in the figure, the hyperspectral image and machine learning collaborative lithology automatic classification system comprises: Figure 2

[0093] The acquisition module 100 is used for collecting training sample polygons, test samples and rock samples in the field;

[0094] ​The spectral measurement module 200 is configured to measure a typical spectral curve of the rock sample, so as to study the spectral curve characteristics corresponding to each type of lithology.

[0095] The dimension reduction processing module 300 is configured to perform two-stage dimension reduction processing on the hyperspectral image by using the characteristics of the typical spectral curve.

[0096] The classification modeling module 400 is configured to construct a classification model by using machine learning, and perform classification on the dimension-reduced hyperspectral image by using the classification model, and generate a lithology classification result map.

[0097] The precision verification module 500 is configured to verify and evaluate the classification precision of the lithology classification result map.

[0098] The specific definition of the lithology automatic classification system based on the cooperation of the hyperspectral image and the machine learning can refer to the definition of the lithology automatic classification method based on the cooperation of the hyperspectral image and the machine learning in the foregoing, and will not be described herein. The modules in the lithology automatic classification system based on the cooperation of the hyperspectral image and the machine learning can be realized by software, hardware, or a combination thereof. The modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so as to be called and executed by the processor to perform the operations corresponding to the modules.

[0099] In one embodiment, a computer device is provided, which can be a server, and an internal structure diagram of the computer device can be as shown in FIG. 8. Figure 4 The computer device includes a processor, a memory, a network interface, and a database connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is configured to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement a lithology automatic classification method based on the cooperation of a hyperspectral image and machine learning.

[0100] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above functional units and modules is exemplified, and in actual applications, the above functions can be completed by different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above.

[0101] The above-described embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A lithology automatic classification method based on hyperspectral image and machine learning, characterized in that, The application relates to a method for classifying lithology based on high-spectral remote sensing images. The method comprises the following steps: Collecting training sample patches, test samples and rock samples in the field; Measuring typical spectral curves of the rock samples to study the spectral curve characteristics corresponding to each type of lithology; Using the spectral curve characteristics to perform two-stage dimension reduction processing on the high-spectral image; Using machine learning to construct a classification model, using the classification model to classify the high-spectral image after dimension reduction, and generating a lithology classification result patch; 2.The lithology automatic classification method of hyperspectral image and machine learning cooperation according to claim 1, characterized in that, Verifying and evaluating the classification accuracy of the lithology classification result patch. The method for collecting training sample patches, test samples and rock samples in the field comprises the following steps: Analyzing the basic geological data of the study area to determine the types and quantities of lithology in the study area; Conducting field reconnaissance, determining the typical distribution area of each type of lithology, and drawing training sample patches; Collecting rock samples with natural weathered surfaces for each type of rock; 3.The lithology automatic classification method of hyperspectral image and machine learning cooperation according to claim 2, characterized in that, Selecting a typical area and conducting geological lithology mapping as test samples to draw the accurate distribution boundary of the rock mass. The method for measuring the typical spectral curves of the rock samples to study the spectral curve characteristics corresponding to each type of lithology comprises the following steps: Implementing multiple-angle spectral curve measurement on each rock sample, and recording the spectral curve of each rock sample and the corresponding lithology type during the measurement process; 4. The method of claim 3, wherein, Taking the average value of the spectral curves measured for all samples of each type of rock as the typical spectral curve of the rock in the area. The method for using the spectral curve characteristics to perform two-stage dimension reduction processing on the high-spectral image comprises the following steps: Pretreating the original image, wherein the pretreatment comprises radiation calibration, atmospheric correction and geometric correction to ensure data quality; Combining the terrain and high-resolution remote sensing image to delineate the Quaternary covering area of the study area, removing the Quaternary covering area image through mask cutting, and excluding the interference of non-rock targets; Entering the first stage of dimension reduction, comparing the typical spectral curves of each type of rock sample, and manually selecting the spectral frequency band and its adjacent wave band sensitive to lithology classification and identification based on the difference characteristics of the measured spectral curves; 5. The method of automatic lithology classification of hyperspectral imagery in conjunction with machine learning according to claim 4, characterized in that, Entering the second stage of dimension reduction, and using the Pearson correlation coefficient (Pearson) method for dynamic threshold dimension reduction based on the manual preliminary selection. The specific operation process of the second stage of dimension reduction comprises: Comparing the correlation coefficients between the wave band data, calculating the correlation between each wave band and other wave bands, and selecting the wave bands with smaller correlation for use; Cyclically selecting wave bands, calculating the average correlation between the remaining wave bands and the selected wave bands in each cycle, and selecting the wave band with the lowest correlation to join the selected wave band list; 6. The method of automatic lithology classification of hyperspectral imagery in conjunction with machine learning according to claim 5, characterized in that, Until the cyclic selection of wave bands reaches the maximum preset number of reserved wave bands or the correlation of all remaining wave bands is higher than a preset threshold of a certain correlation, the selection of wave bands is stopped, and the wave bands after dimension reduction are combined into a new GeoTIFF file. The method for using machine learning to construct a classification model, using the classification model to classify the high-spectral image after dimension reduction, and generating a lithology classification result patch comprises the following steps: Using the training sample patches drawn in the field to extract the corresponding remote sensing image spectral values to construct classification samples; Using the classification model to classify the high-spectral image after dimension reduction, and generating a lithology classification result patch; Random forest (RF), support vector machine (SVM) or XGboost algorithm is selected to model the classification samples, and the hyperparameters of the model are adjusted and optimized through grid search combined with cross-validation method; The optimal hyperparameter combination is used to train the classification model, and the classification model is used to perform end-to-end inference on the reduced hyperspectral image to generate a lithology classification result patch.

7. The method of claim 1, wherein, The classification accuracy of the lithology classification result patch is verified and evaluated, including the following steps: Comparing the result patch of the geological lithology mapping and the lithology classification result patch; The overall classification accuracy and Kappa coefficient are calculated through the confusion matrix, which is used to compare and evaluate the lithology classification results of different machine learning algorithms, and based on the quantitative indicators such as the overall classification accuracy and Kappa coefficient, the classification errors of different lithology types are evaluated; Spectral feature analysis is performed on the misclassified area to trace the error source; The performance of different machine learning algorithms is compared, and the optimal model is selected for regional promotion to form a complete lithology classification technology closed loop.

8. A lithology automatic classification system based on hyperspectral image and machine learning, characterized in that, It includes: A collection module for collecting training sample patches, test samples and rock samples in the field; A spectral measurement module for measuring the typical spectral curve of the rock sample, which is used to study the spectral curve characteristics corresponding to each type of lithology; A dimension reduction processing module for performing two-stage dimension reduction processing on the hyperspectral image using the characteristics of the typical spectral curve; A classification modeling module for constructing a classification model using machine learning, and using the classification model to classify the reduced hyperspectral image and generate a lithology classification result patch; An accuracy verification module for verifying and evaluating the classification accuracy of the lithology classification result patch.

Images