Lithology classification method based on spectrum space residual network

Abstract

The invention relates to a lithology classification method based on a spectrum space residual network, and the method comprises the following steps: collecting multi-source geoscience data of a detected region, carrying out the data processing of the multi-source geoscience data, and forming a to-be-classified image; lithologic sample positions and lithologic category codes are marked on the to-be-classified images; lithologic samples are extracted from the to-be-classified images, and the independence of the samples is checked; establishing a lithology classification model based on a spectrum space residual network architecture, setting model hyper-parameters, training the model, analyzing the accuracy and loss value index of the model, and checking the robustness of the model; lithology information in the to-be-classified image is extracted, and lithology unit boundaries are delineated and corrected; and splicing the lithology unit boundary in the classification result grid image with the boundary of the non-rock bare area, and compiling a complete lithology classification result map of the detected area. According to the method, accurate and efficient lithology classification is realized, and geological personnel can be helped to better carry out regional geological survey.

Description

Technical Field

[0001] The invention relates to a method for multi-source data-assisted regional geological survey, in particular to a lithology classification method based on spectrum space residual network. Background Art

[0002] Existing lithology classification methods can be roughly divided into three categories: unsupervised classification, semi-supervised classification and supervised classification. Among them, the unsupervised classification method does not set samples in advance, and mainly implements lithology classification based on the characteristic differences within the data. Representative algorithms include K-Means, ISODATA, etc. The semi-supervised classification method combines the internal difference characteristics of the data with expert experience to set a small number of samples, and implements lithology classification based on this. Representative algorithms include self-training method, semi-supervised support vector machine method, etc. The supervised classification method first sets a certain number of samples based on expert experience, and implements lithology classification based on these samples. Representative algorithms include maximum likelihood method, neural network method, etc.

[0003] Most of the existing lithology classification methods are based on mathematical statistics or shallow machine learning. The algorithm principle is relatively simple, and the classification model has a weak ability to fit complex lithology characteristics. Therefore, the current lithology classification work in areas with complex geological conditions still has the problem of low accuracy. Summary of the invention

[0004] The purpose of the present invention is to provide a lithology classification method based on a spectrum spatial residual network to solve the problem of low classification accuracy in existing lithology classification methods.

[0005] The object of the present invention is achieved in that:

[0006] A lithology classification method based on spectrum spatial residual network includes the following steps:

[0007] S1. Data collection: Collect multi-source geological data including geological data, aeromagnetic and aerial radio data and hyperspectral remote sensing images in the surveyed area.

[0008] S2. Process data: perform routine processing on the collected multi-source geological data, then select data channels, normalize the image data to unify the format, projection, grid density and spatial range of the image data, and combine the processed images of different types into a multi-channel image; circle the non-rock exposed areas including water bodies, alluvial deposits, artificial buildings and engineering distribution areas on the multi-channel image, and remove the data of the non-rock exposed areas from the multi-channel image to form an image to be classified of the measured area.

[0009] S3. Prepare a lithology sample set: establish discriminant marks for various types of lithology on the image to be classified, and mark the position of the lithology samples and their lithology category codes on the image to be classified; extract lithology samples from the image to be classified according to the marked lithology sample positions, and test the independence of the samples; perform sorting operations on the lithology samples, including sample shuffling, three-dimensional sample expansion and sample data normalization, and divide the sorted lithology samples into training set, validation set and test set in proportion to form a lithology sample set.

[0010] S4. Construct a lithology classification model: A lithology classification model is established based on the spectrum-space residual network architecture. The input port, convolution kernel size, number of convolution layers and output port of the lithology classification model are adjusted according to the sample structure, and the model hyperparameters are initialized.

[0011] S5. Training model: Use the training set and validation set to train the lithology classification model, and use the test set to test the robustness of the model; comprehensively evaluate the classification performance of the lithology classification model based on the accuracy and loss value indicators of the model on the training set and validation set, and the confusion matrix classification accuracy on the test set. If the performance analysis results do not meet the standards, adjust the model hyperparameters item by item, and repeat this step; until the performance analysis results meet the standards, determine the model hyperparameters at this time, and save the model.

[0012] S6. Output lithology classification results: Use the Gdal library in Python to load the image to be classified, extract the lithology information and geographic parameter information including grid size, number of rows and columns, and spatial projection in the image to be classified; normalize the extracted image data information and then input it into the lithology classification model, which outputs the lithology classification results expressed in Roman numerals; use the Gdal library to convert the output lithology classification results into a classification result raster image with the same geographic parameter information as the image to be classified; import the classification result raster image into ArcGIS software to outline and correct the lithology unit boundaries.

[0013] S7. Prepare a lithology classification result map: splice together the boundaries of the lithology units in the classification result raster image with the boundaries of the non-rock exposed areas to prepare a complete lithology classification result map of the measured area; if there are areas in the lithology classification result map of the measured area that are inconsistent with the existing geological knowledge or areas where the lithology classification results are questionable, conduct a field survey and revise the lithology classification result map of the measured area based on the survey results.

[0014] The present invention is based on the systematic collection of multi-source geological data of the measured area, firstly forms an image to be classified according to the processed data, then uses the image to be classified to make a lithology sample set, and then builds a lithology classification model based on the spectrum space residual network, trains the model, saves the model with the best classification performance, uses the best model to output the lithology classification result grid image, and draws the lithology unit boundary, and finally, comprehensively compiles the lithology classification result map of the measured area. The present invention realizes fast, accurate, intelligent and efficient lithology classification under complex geological conditions, overcomes the problem of low accuracy of existing lithology classification methods, and is conducive to assisting geological personnel to better carry out regional geological surveys. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flowchart of the lithology classification method of the present invention.

[0016] Figure 2 It is a structural diagram of the lithology classification model. DETAILED DESCRIPTION

[0017] Limitations of the present invention: The lithology classification involved in the present invention is limited to image data.

[0018] The present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0019] like Figure 1 As shown, the lithology classification method of the present invention comprises the following steps:

[0020] Step 1: Collect data.

[0021] Collect multi-source geological data including geological data, aeromagnetic and aerial radio data and hyperspectral remote sensing images in the surveyed area. In this embodiment, the Revolutionary Valley area of ​​Longshou Mountain is selected as the surveyed area, and the geological data, aeromagnetic and aerial radio data and hyperspectral remote sensing images in the area are collected. In theory, the more diverse the data types, the richer the physical property information provided. The geological data collected this time are 1:50,000 regional geological reports and related maps, the aeromagnetic and aerial radio data are 1:50,000 high-precision aeromagnetic and aerial radio data, and the hyperspectral remote sensing images are domestic resource-one 02D (ZY1-02D) satellite hyperspectral images. It should be noted that the geological and remote sensing data used in this embodiment are all unprocessed raw data, while the aeromagnetic and aerial radio data are standard database files that have been pre-processed in the field.

[0022] Step 2: Process the data.

[0023] 1. Perform routine processing on the collected multi-source geological data. Routine processing operations include: vectorization of geological maps for geological data; cartographic conversion of aeromagnetic and aerial radio data; radiation correction, atmospheric correction, orthorectification, geometric precision correction and abnormal band detection for hyperspectral remote sensing images.

[0024] In this embodiment, Mapgis and ArcGIS software are used to vectorize geological maps. According to the remote sensing image preprocessing technical regulations, ENVI software is used to perform radiation correction, atmospheric correction, orthorectification, geometric precision correction and abnormal band detection on ZY1-02D ​​satellite hyperspectral images. According to the aeromagnetic and airborne data processing procedures, Geosoft software was used to carry out mapping conversion processing on 1:50,000 high-precision aeromagnetic and airborne data. Seven conversion maps were produced based on the aeromagnetic data, including aeromagnetic △T polar plane, aeromagnetic △T polar upward extension plane, aeromagnetic △T polar horizontal gradient mode plane, aeromagnetic △T polar vertical first-order derivative plane, and aeromagnetic △T polar oblique derivative plane; seven conversion maps were produced based on the aeromagnetic data, including aeromagnetic total volume contour line plane, aeromagnetic potassium content contour line plane, aeromagnetic uranium content contour line plane, aeromagnetic thorium content contour line plane, aeromagnetic uranium / potassium contour line plane, aeromagnetic uranium / thorium contour line plane, and aeromagnetic thorium / potassium contour line plane.

[0025] 2. Optimize the data channel and normalize the image data to unify the format, projection, grid density and spatial range of the image data; synthesize the processed images of different types into a multi-channel image; circle the non-rock exposed areas including water bodies, alluvial deposits, artificial buildings and engineering distribution areas on the multi-channel image within the measured area, and remove the data of the non-rock exposed areas on the multi-channel image, so as to form an image to be classified in the measured area.

[0026] Hyperspectral remote sensing images have many bands, and there is a strong correlation between the information of different bands. There may also be a certain correlation between the various conversion maps of aeromagnetic and aerial radiography. In order to reduce the interference of correlation between data on the analysis results, this embodiment combines the Optimum Index Factor (OIF) and Correlation Analysis to optimize the data channels. In the end, hyperspectral remote sensing data contributed 52 data channels, aeromagnetic conversion maps contributed 6 data channels, and aerial radiography conversion maps contributed 6 data channels, totaling 64 data channels.

[0027] The parameters of hyperspectral remote sensing and aeromagnetic and aerial release data were standardized by using ENVI, ArcGIS, and Geosoft software. The data formats were all raster image files, the spatial projections were all WGS_1984_UTM_Zone_47N, the grid size (i.e., grid size) was 30 meters × 30 meters, and the spatial range was uniformly adjusted to be consistent with the measured area. Finally, the composite bands tool of ArcGIS software was used to superimpose the hyperspectral remote sensing data channels and aeromagnetic and aerial release data channels into a total multi-channel image, with a total of 64 channels.

[0028] Step 3: Prepare a lithology sample set.

[0029] 1. Establish the discriminant marks of various lithologies on the image to be classified, and mark the position of the lithology samples and their lithology category codes on the image to be classified. The specific operation method of marking the position of the lithology samples on the image to be classified is: according to the discriminant marks of various lithologies on the image to be classified, use ArcGIS software to mark the position of the lithology samples on the image to be classified, save the sample position as a vector surface file, and add an attribute field to store the lithology category code in the vector surface file, and then convert the vector file into a raster form and record it as a label image. The lithology category code is represented by Roman numerals starting from "I", each numeral represents a lithology category, and the lithology category where there is no marked sample on the image to be classified is "0".

[0030] The way to establish the distinguishing marks of various lithologies on the image to be classified is to carry out field surveys appropriately according to the existing geological data of the measured area, summarize the main rock types in the measured area, and establish the distinguishing marks of various lithologies on the image to be classified on this basis. If necessary, typical rocks should also be collected during the survey for rock and mineral identification experiments to assist the staff in determining the type of the target lithology.

[0031] In this embodiment, there are six main types of lithology in the measured area: alkaline complex, granite, phyllite, schist, marble and siliceous limestone. From the perspective of spatial distribution, alkaline complex and granite are concentrated and have a large exposed area, followed by phyllite, schist and marble, and siliceous limestone is constrained by fault structures and is distributed in a belt-like manner with a small area. Overall, the geological conditions in the measured area are relatively complex.

[0032] In terms of marking the sample positions, the six rock types, namely alkaline melange, granite, phyllite, schist, marble and siliceous limestone, were coded with Roman numerals I to X, and counted according to the number of pixels. The number of rock type samples initially marked were 278, 512, 389, 123, 278, and 240, respectively, totaling 1,820.

[0033] 2. According to the marked lithological sample positions, extract lithological samples from the image to be classified, and test the independence of the lithological samples to ensure that the sample quality meets the standards. The specific operation method is: use ENVI software to calculate the transformation separation of the marked lithological samples, and test the independence of the lithological samples based on the calculation results; if the independence of the test does not meet the standards, reselect the lithological samples and repeat the independence test until the independence of the lithological samples all meet the standards. Specifically, use the Compute ROI Separability tool in ENVI software to calculate the transformed divergence between the marked lithological samples. The results show that the transformed divergence is greater than 1.9, indicating that the independence of the samples meets the requirements.

[0034] 3. Perform operations such as sample shuffling, three-dimensional sample expansion, and sample data normalization on the lithology samples. Through sample division, a lithology sample set consisting of a training set, a validation set, and a test set is formed.

[0035] 3.1 Sample shuffling operation for lithological samples. Since there may be a certain spatial correlation between adjacent samples when marking samples, in order to avoid interference of spatial correlation on the analysis results, the order of lithological samples is shuffled. The specific operation method for shuffling lithological samples is: use the shuffle() function of the random module in Python to shuffle the order of the positions of lithological samples. Specifically, use the Gdal library in Python to load the label image and convert it into an array form using the Numpy library. The values ​​in the array correspond one-to-one to the values ​​on the label image. Traverse the points in the array whose values ​​are not 0, save their position information in the form of a list, and then use the shuffle() function of the random module to shuffle the list. In this way, the order of samples is shuffled.

[0036] 3.2 The operation of expanding and selecting three-dimensional samples of lithology samples is to make full use of the spectrum and spatial information of the selected data to achieve better lithology classification results. Specifically, a three-dimensional lithology sample is made according to the three-dimensional data structure of "one-dimensional spectrum × two-dimensional space". On the basis of sample disorder, it is assumed that the new position of a sample is A, and the pixel at A is centered, so that all data channels are covered in the spectrum and the appropriate size is expanded in space as a three-dimensional range, and all data within this range are recorded as the data of the sample. In other words, the sample data structure can be recorded as C×L×L, where C represents the number of spectrum channels and L represents the spatial size. Specifically, the image to be classified is loaded using the Gdal library in Python, and converted into an array form using the Numpy library. The new position of the sample is used to match and extract the corresponding information in the array. According to the data characteristics and hardware and software facilities, the data structure of a single lithology sample in this embodiment is set to 64×5×5, that is, the spectrum dimension has 64 data channels and the spatial dimension is 5×5 grids in size. Considering that the total number of samples is 1820, the structure of the lithology sample set prepared in this embodiment can be recorded as 1820×64×5×5.

[0037] 3.3 The sample data normalization operation of lithology samples is because the dimensions and ranges of the data of each channel often vary greatly, which usually has a great impact on the analysis efficiency and accuracy. Therefore, the sample data should be normalized. Through normalization, the range of the sample data is uniformly reduced to a reasonable range and the dimension difference is eliminated, which significantly improves the analysis effect.

[0038] To ensure the analysis effect, the present invention recommends the use of the maximum-minimum normalization method, which can convert the data value range into the [0,1] interval. The calculation formula is:

[0039]

[0040] Where C is a channel in the sample data, Cx is the initial data of the channel, Cmin is the minimum value in the initial data, Cmax is the maximum value in the initial data, and Cx * is the result value after normalization.

[0041] The specific method of sample data normalization is to use the max() function and min() function of the Numpy library in Python to calculate the maximum and minimum values ​​of a single data channel, and then perform maximum-minimum value normalization on this basis. Finally, the values ​​are converted to the [0,1] interval.

[0042] 3.4 The sorted lithology samples are divided into three parts: training set, validation set and test set according to a certain ratio. The training set and validation set are used to train the lithology classification model, and the test set is used to test the robustness of the model. In this embodiment, the split() function in Python is used to divide the samples according to the ratio of "training set: validation set: test set = 6:2:2". The specific number of lithology samples in the training set, validation set and test set is 1092, 364 and 364 respectively. At this point, the process of making the lithology sample set is completed, and the lithology sample set is formed.

[0043] Step 4: Construct a lithology classification model.

[0044] A lithology classification model is established based on the spectrum-space residual network architecture. The input port, convolution kernel size, number of convolution layers and output port of the lithology classification model are adjusted according to the sample structure, and the hyperparameters of the model are initialized. The core components of the classic spectrum-space residual network structure are three: three-dimensional convolution block, spectrum residual block and space residual block. The present invention constructs a lithology classification model based on the Pytorch development framework.

[0045] like Figure 2 As shown, the lithology classification model constructed by the present invention refers to the classic spectrum space residual network structure and is appropriately adjusted according to the structure of the lithology sample. The constructed lithology classification model includes an input layer, a convolution layer, two spectrum residual blocks, two convolution layers, two space residual blocks, a pooling layer and a fully connected layer from front to back; its specific structure is: the structure of the input port is 5×5×64, the size of the spectrum convolution kernel is 1×1×7×28, the size of the space convolution kernel is 3×3×28, and the structure of the output port is 1×1×6; at the same time, the model hyperparameters including learning rate, batch size, number of iterations, optimizer, activation function, and loss function are initialized and set according to conventional methods.

[0046] Step 5: Train the model.

[0047] The lithology classification model is trained using the sample data of the training set and the validation set, and the robustness of the model is tested using the test set. The classification performance of the lithology classification model is comprehensively evaluated based on the accuracy and loss value indicators of the model on the training set and the validation set, as well as the confusion matrix classification accuracy on the test set. If the performance analysis results do not meet the standards, the model hyperparameters are adjusted item by item, and this step is repeated until the performance analysis results meet the standards. The model hyperparameter combination at this time is determined, and the model with the best performance is saved.

[0048] The operation method for testing the robustness of the model is: input the samples of the test set into the lithology classification model, and then extract the sample data and label part in the test set; use the sample data to perform lithology classification processing, and calculate the classification accuracy of the lithology classification model for the lithology samples in the test set based on the classification results and the corresponding actual labels, and test the robustness of the model based on the classification accuracy; if the robustness of the model does not meet the standard, reselect the lithology samples and repeat steps three to five until the robustness of the model meets the standard.

[0049] A confusion matrix is ​​established based on the classification results and the corresponding actual labels, and the classification accuracy of the lithology classification model on the test set is evaluated based on the confusion matrix. The evaluation indicators include accuracy, precision, recall and F1 score.

[0050] 1. The calculation formula of accuracy index is:

[0051]

[0052] Among them, ACC represents accuracy, TP represents true positive examples, TN represents true negative examples, FP represents false positive examples, and FN represents false negative examples.

[0053] 2. The calculation formula of the precision rate indicator is:

[0054]

[0055] Among them, PPV represents precision, TP represents true positive, and FP represents false positive.

[0056] 3. The calculation formula of recall rate is:

[0057]

[0058] Among them, TRP represents recall rate, TP represents true positive examples, and FN represents false negative examples.

[0059] 4. The calculation formula of F1 score index is:

[0060]

[0061] Among them, F1 represents F1 score, PPV represents precision, and TRP represents recall.

[0062] Dynamic adjustments were made to the six key hyperparameters of the model, namely, learning rate, batch size, number of iterations, optimizer, activation function, and loss function. The adjustment method was the control variable method, that is, only one hyperparameter was changed each time, and the parameter combination with the best classification performance was compared and selected for the next round of adjustments. The classification performance of the model was comprehensively evaluated based on the accuracy and loss value indicators of the model on the training set and validation set, as well as the confusion matrix classification accuracy on the test set. Generally, the higher the accuracy of the model on the training set and validation set, the smaller the fluctuation, the lower the loss value, and the faster the convergence, the better the training effect of the model. The higher the confusion matrix classification accuracy of the model on the test set, the better the robustness of the model.

[0063] After adjusting the parameters, this embodiment determined that the best hyperparameter combination is: learning rate = 0.001, batch size = 32, number of iterations = 200, optimizer = RMSProp, activation function = ReLU, loss function = Cross Entropy. The best classification accuracy of the model for the training set is 97.50%, the best classification accuracy of the model for the validation set is 93.93%, the overall accuracy of the model for the lithology classification of the test set is 92.31%, the precision rate of the six lithologies is not less than 88.00%, the recall rate is not less than 90.24%, and the F1 score is not less than 0.89 (see Table 1). Finally, use the save() function of the Torch library in Python to save the hyperparameters of the best model as a model file with the suffix pth.

[0064] Table 1. Lithology classification accuracy calculation results

[0065]

[0066] Step six: output the lithology classification results.

[0067] The Gdal library in Python is used to load the image to be classified, and the lithology information and geographic parameter information including grid size, number of rows and columns, and spatial projection in the image to be classified are extracted; the extracted image data information is normalized and then input into the lithology classification model, which outputs the lithology classification results expressed in Roman numerals. The Gdal library is used to convert the output lithology classification results into a classification result raster image with the same geographic parameter information as the image to be classified. The classification result raster image is imported into ArcGIS software to outline and correct the lithology unit boundaries.

[0068] In this embodiment, the data in the image to be classified is input into the best lithology classification model to obtain the classification result raster image of six lithologies in the measured area, and then the accurate lithology unit boundary is outlined in ArcGIS software.

[0069] Step seven, make a lithology classification result map.

[0070] The boundaries of the lithologic units in the classification result grid image are spliced ​​together with the boundaries of the non-rock exposed areas to compile a complete lithologic classification result map of the measured area. If there are areas in the lithologic classification result map of the measured area that are inconsistent with the existing geological knowledge or areas where the lithologic classification results are questionable, a field survey is conducted and the lithologic classification result map of the measured area is revised based on the survey results.

[0071] The lithology classification method of the present invention is compared and verified with the lithology classification method based on the maximum likelihood method, that is, the sample data of the test set in this embodiment is subjected to lithology classification processing, and the accuracy values ​​are shown in Table 2.

[0072] Table 2. Comparison of classification accuracy of test set sample data

[0073]

[0074] From the comparative data in Table 2, it can be seen that in terms of accuracy, the value of the lithology classification method of the present invention is 92.31%, and the value of the maximum likelihood method is 87.91%; in terms of the F1 scores of the six lithologies, the calculation results of the lithology classification method of the present invention are all greater than the calculation results of the maximum likelihood method. It can be seen that compared with the existing lithology classification method, the lithology classification method of the present invention has higher classification accuracy and better use effect.

Claims

1. A lithology classification method based on spectrum spatial residual network, characterized in that: The following steps are involved: S1. Data collection: Collect multi-source geoscientific data including geoscientific data, aeromagnetic and aerial radio data and hyperspectral remote sensing images in the surveyed area; S2. Data processing: Perform routine processing on the collected multi-source geological data, then select data channels, normalize the image data to unify the format, projection, grid density and spatial range of the image data, and combine the processed images of different types into a multi-channel image; delineate non-rock exposed areas including water bodies, alluvial deposits, artificial buildings and engineering distribution areas on the multi-channel image, and remove the data of the non-rock exposed areas from the multi-channel image to form an image to be classified of the measured area; S3. Prepare a lithology sample set: establish identification marks for various lithologies on the image to be classified, and mark the location of the lithology sample and its lithology category code on the image to be classified; According to the marked lithology sample positions, lithology samples are extracted from the images to be classified, and the independence of the samples is tested; the lithology samples are sorted, including sample shuffling, three-dimensional sample expansion, and sample data normalization, and the sorted lithology samples are divided into a training set, a validation set, and a test set in proportion to form a lithology sample set; S4. Constructing a lithology classification model: Building a lithology classification model based on the spectrum spatial residual network architecture, adjusting the input port, convolution kernel size, number of convolution layers and output port of the lithology classification model according to the sample structure, and initializing the hyperparameters of the model; S5. Training model: Use the training set and validation set to train the lithology classification model, and use the test set to test the robustness of the model; comprehensively evaluate the classification performance of the lithology classification model based on the accuracy and loss value indicators of the model on the training set and validation set and the confusion matrix classification accuracy on the test set. If the performance analysis result does not meet the standard, adjust the model hyperparameters item by item, and repeat this step until the performance analysis result meets the standard, determine the model hyperparameters at this time, and save the model; S6. Output lithology classification results: Use the Gdal library in Python to load the image to be classified, extract the lithology information and geographic parameter information including grid size, number of rows and columns, and spatial projection in the image to be classified; normalize the extracted image data information, and then input it into the lithology classification model, which outputs the lithology classification results expressed in Roman numerals; Using the Gdal library, the output lithology classification results are converted into a classification result raster image with the same geographic parameter information as the image to be classified; the classification result raster image is imported into ArcGIS software to outline and correct the lithology unit boundaries; S7. Prepare a lithology classification result map: splice together the boundaries of the lithology units in the classification result raster image with the boundaries of the non-rock exposed areas to prepare a complete lithology classification result map of the measured area; if there are areas in the lithology classification result map of the measured area that are inconsistent with the existing geological knowledge or areas where the lithology classification results are questionable, conduct a field survey and revise the lithology classification result map of the measured area based on the survey results.

2. The lithology classification method based on spectrum spatial residual network according to claim 1 is characterized in that: The conventional processing operations on multi-source geological data in step S2 include: vectorizing geological maps of geological data; performing cartographic conversion processing on aeromagnetic and aerial radio data; and performing radiation correction, atmospheric correction, orthorectification, geometric precision correction and abnormal band detection processing on hyperspectral remote sensing images.

3. The lithology classification method based on spectrum spatial residual network according to claim 1 is characterized in that: The operation method of marking the position of the lithology sample on the image to be classified in step S3 is: according to the discrimination marks of various lithologies on the image to be classified, the position of the lithology sample is marked on the image to be classified using ArcGIS software, the sample position is saved as a vector surface file, and an attribute field for storing the lithology category code is added to the vector surface file, and then the vector file is converted into a raster form and recorded as a label image; the lithology category code is represented by Roman numerals starting from "I", each numeral represents a lithology category, and the lithology category where there is no marked sample on the image to be classified is "0".

4. The lithology classification method based on spectrum spatial residual network according to claim 1 is characterized in that: The operation method of testing the independence of samples in step S3 is: using ENVI software to calculate the conversion separation of the marked lithological samples, and analyzing the independence of the lithological samples based on the calculation results; if the analysis results do not meet the standards, reselect the lithological samples and repeat the independence test until the independence of the lithological samples all meet the standards.

5. The lithology classification method based on spectrum spatial residual network according to claim 1 is characterized in that: The method of shuffling the lithology samples in step S3 is to use the shuffle() function of the random module in Python to shuffle the order of the lithology samples.

6. The lithology classification method based on spectrum spatial residual network according to claim 1 is characterized in that: The structure of the lithology classification model constructed in step S4 includes, from front to back, an input layer, a convolutional layer, two spectral residual blocks, two convolutional layers, two spatial residual blocks, a pooling layer and a fully connected layer; the model hyperparameters initialized include learning rate, batch size, number of iterations, optimizer, activation function and loss function.

7. The lithology classification method based on spectrum spatial residual network according to claim 1 is characterized in that: The operation method for testing the robustness of the model in step S5 is: input the samples of the test set into the lithology classification model, and then extract the sample data and label part in the test set; use the sample data to perform lithology classification processing, and calculate the classification accuracy of the lithology classification model for the lithology samples in the test set according to the classification results and the corresponding actual labels, and test the robustness of the model based on the classification accuracy; if the robustness of the model does not meet the standard, reselect the lithology samples and repeat steps S3 to S5 until the robustness of the model meets the standard.

Images