Intelligent lithology identification method based on generative adversarial network

Abstract

The invention belongs to the technical field of hyperspectral remote sensing, and particularly relates to a lithology intelligent identification method based on a generative adversarial network, and the method comprises the steps: 1, obtaining a high-spatial-resolution remote sensing image of a research region, and carrying out the field investigation; 2, preprocessing the hyperspectral remote sensing data, and removing abnormal wavebands; 3, according to lithology information combining field investigation and geological data, image sample delineation is carried out on the preprocessed hyperspectral image, and an original training set is obtained; 4, constructing a deep convolutional generative adversarial network DCGAN comprising a generative network G and a discriminant network D; 5, introducing a convolution operation and a self-attention mechanism, and optimizing a generative adversarial network structure; 6, training and quality evaluation of the generated samples are carried out, and a newly generated sample set is obtained; and step 7, constructing a one-dimensional + two-dimensional convolutional neural network model based on graph-spectrum joint features. According to the method, expansion and optimization of the sample data are realized, and the lithology identification precision is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of hyperspectral remote sensing, and in particular relates to an intelligent lithology identification method based on a generative adversarial network. Background Art

[0002] Hyperspectral remote sensing technology provides important information for basic geological surveys and mineral resource exploration because it can obtain spectra of up to hundreds of bands in a macro, fast and large-scale manner. Among them, the use of hyperspectral remote sensing data for lithology identification is the main application direction of hyperspectral remote sensing in the field of geological exploration. Although traditional hyperspectral remote sensing methods have achieved good results in mineral identification, there are still certain bottlenecks in lithology identification. In recent years, machine learning methods have been increasingly introduced into remote sensing geological exploration. This type of data-driven method can automatically learn the relationship between remote sensing data and the required features. However, the recognition accuracy of deep learning depends on the input of a large amount of training data. The types of objects in the currently public hyperspectral image data sets are mostly farmland, woodland, water bodies and artificial buildings. They cannot be directly input as training sets for lithology identification. At the same time, since the research areas for lithology identification are mostly located in the northwest region with large terrain cuts, low vegetation coverage, wide bedrock exposure area, and sparse population, the number of lithology point information obtained solely by relying on field surveys is difficult to meet the requirements of machine learning training; and simply using relevant geological data such as geological maps to delineate training data set samples is limited by the scale of geological maps and it is difficult to obtain accurate lithology data information. Therefore, the number of training data sets greatly limits the improvement of lithology identification accuracy.

[0003] Generative adversarial networks (GANs) are a type of deep learning network that can obtain the implicit feature distribution of existing data and synthesize new data samples in an unsupervised learning manner. A series of derivative models of GAN are mostly used in fields such as data enhancement, image classification and image synthesis, but the image classification involved is mostly for optical image data, and is rarely used in hyperspectral data classification, especially hyperspectral data lithology classification.

[0004] Therefore, it is urgent to develop an intelligent lithology identification method based on generative adversarial networks to overcome the defects of existing lithology identification methods. Summary of the invention

[0005] The purpose of the present invention is to provide a method for intelligent lithology identification based on a generative adversarial network, which realizes the expansion and optimization of sample data and effectively improves the accuracy of lithology identification.

[0006] The technical solution to achieve the purpose of the present invention is:

[0007] A method for intelligent lithology identification based on a generative adversarial network, the method comprising:

[0008] Step 1: Obtain high spatial resolution remote sensing images of the study area and conduct field surveys;

[0009] Step 2: Preprocess the hyperspectral remote sensing data and remove abnormal bands;

[0010] Step 3: Based on the lithology information obtained from field surveys and geological data, image samples are delineated on the pre-processed hyperspectral images to obtain the original training set;

[0011] Step 4: Construct a deep convolutional generative adversarial network DCGAN including a generator network G and a discriminator network D;

[0012] Step 5: Introduce convolution operation and self-attention mechanism to optimize the generative adversarial network structure;

[0013] Step 6: Generate sample training and quality evaluation to obtain a new generated sample set;

[0014] Step 7: Construct a one-dimensional + two-dimensional convolutional neural network model based on graph-spectrum joint features.

[0015] The step one is specifically as follows: obtaining high spatial resolution remote sensing images of the study area, circling the bare rock locations that can be obtained by visual interpretation on the images, designing field verification routes and verification points, conducting field surveys, verifying the lithology information of the outcrop locations, and if there are inaccessible locations in the designed route, supplementing the lithology information with geological data.

[0016] The hyperspectral remote sensing data preprocessing in step 2 includes: radiation calibration, atmospheric correction, and image clipping; the abnormal bands include 1400nm, 1900nm, and 2400nm.

[0017] The step three is specifically as follows: according to the lithology information combined with the field survey and the geological map, the sample category is determined, and the image sample of each type of lithology is circled on the pre-processed hyperspectral image to obtain the lithology image and spectral data; the size of the circled image sample is adjusted so that the size of the circled image sample remains consistent to adapt to the input of the network; the spectral data corresponding to the circled image sample is normalized, and the obtained image and spectral training set is the original training set.

[0018] The fourth step comprises:

[0019] Step 4.1: Input data: Use random noise as the input of the generator; use the actual hyperspectral remote sensing image as the input of the discriminator as a real sample;

[0020] Step 4.2: Generate "fake" samples: The generator generates "fake" sample images through multi-layer convolution and deconvolution operations using a convolutional neural network (CNN). The size and spectral characteristics of the "fake" sample images are close to those of real samples.

[0021] Step 4.3: Distinguish true and false samples: The discriminator compares the "false" sample with the real sample, calculates its authenticity score, and outputs whether the sample is a real sample;

[0022] Step 4.4: Network optimization: The generator and discriminator gradually update their respective parameters during the confrontation process. The generator optimizes its generation ability by minimizing the discrimination error of the discriminator, while the discriminator optimizes its discrimination ability by maximizing its distinction between "fake" samples and "real" samples, so that the generated "fake" samples are getting closer and closer to the real data distribution.

[0023] The step five comprises:

[0024] Step 5.1: Generator optimization: Add more convolutional layers and batch normalization layers on the basis of the original DCGAN. Each convolution operation gradually enlarges the spatial resolution of the feature map through the deconvolution technology, and finally outputs a spectral data image of size 128×128;

[0025] Step 5.2: The generator generates "fake" samples: The input of the generator is a uniformly distributed random noise vector. Through layer-by-layer deconvolution operations, a hyperspectral data image with a shape of 128×128 is finally generated, which is the original feature map. Each layer of the generator is batch normalized to ensure the stability of the model training process. At the same time, the generator output layer uses the Tanh activation function to compress the generated data to the range of [-1,1] to meet the standard format of hyperspectral data.

[0026] Step 5.3: Discriminator optimization: The structure of the discriminator is optimized accordingly. A transposed convolution layer is added to the original DCGAN, and a batch normalization layer is introduced in the middle layer. The last layer of the discriminator uses the Sigmoid activation function to output a probability value between 0 and 1, indicating the probability that the sample is real data. The other convolutional layers use the Leaky ReLU activation function to help the network better handle negative inputs and prevent the gradient vanishing problem.

[0027] Step 5.4: The discriminator determines true or false: The discriminator receives the "false" samples from the generator and the actual hyperspectral samples, performs convolution and pooling operations layer by layer, and outputs the true or false judgment value;

[0028] Step 5.5: Introduce the self-attention mechanism and generate the self-attention feature map: Introduce the self-attention mechanism module in the network structure; use two 1×1 convolutions to perform linear transformation on the convolution feature map and convert it into a matrix form; calculate the attention weight of the feature map, that is, the similarity matrix, and obtain the normalized attention map through the Softmax function; multiply the original feature map in step 5.2 with the attention map and weight them to generate the final self-attention feature map.

[0029] The step six is ​​specifically as follows: the generator and the discriminator are continuously optimized during the training process, and their respective parameters are iteratively updated to continuously improve the quality of the generated spectral samples; when the generated samples can pass the true judgment of the discriminator and their spectral characteristics are close to the actual samples, it is considered that the generation process has achieved the expected effect, and the generated samples constitute the new generated sample set after training.

[0030] The training of the generated samples in step six is ​​specifically as follows: adopting a strategy of multiple updates of the generator, after each update of the discriminator weight, the generator is updated 5 times, so that the generator can quickly learn the distribution of the spectral data under the feedback of the discriminator and generate more realistic samples.

[0031] The step seven is specifically as follows: the actual sample data in step six and the newly generated samples are combined into a new data set, and then the new data set image is mixed with the original training set image obtained in step three as the input hyperspectral data of CNN; for the spectral branch in the hyperspectral data, multiple alternating one-dimensional convolutional layers and one-dimensional pooling layers are used to extract spectral features; for the image branch, multiple alternating two-dimensional convolutional layers and two-dimensional pooling layers are used to extract image features, and the image features are reconstructed into a one-dimensional tensor; the extracted spectral features and image features are fused and connected, and the joint graph-spectrum features are classified using a fully connected layer, and a one-dimensional + two-dimensional convolutional neural network model based on the joint graph-spectrum features is constructed.

[0032] The activation function used by the one-dimensional convolution layer and the two-dimensional convolution layer is ReLU, and the fusion connection of the extracted spectral features and image features adopts the Concate fusion method.

[0033] The beneficial technical effects of the present invention are:

[0034] 1. The present invention provides an intelligent lithology identification method based on a generative adversarial network. The preliminary training set data is obtained by combining field surveys with geological data, and combined with adversarial network training, the sample data is expanded and optimized, and the accuracy of lithology identification is effectively improved.

[0035] 2. The present invention provides an intelligent lithology identification method based on a generative adversarial network, which increases the effective data volume of lithology label data and effectively improves the accuracy of lithology identification by combining spectral images and textures.

[0036] 3. The present invention provides a method for intelligent lithology identification based on a generative adversarial network, which establishes a generative adversarial network (GAN) and a one-dimensional + two-dimensional convolutional neural network model that combines convolution operation and self-attention mechanism, realizes the expansion of the hyperspectral lithology sample data set, and effectively improves the accuracy of intelligent lithology identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A flow chart of a lithology intelligent identification method based on a generative adversarial network provided by the present invention;

[0038] Figure 2 A structure diagram of a generative adversarial neural network in a lithology intelligent identification method based on a generative adversarial network provided by the present invention;

[0039] Figure 3 This is a 1D+2D CNN structure diagram based on graph-spectrum feature extraction in a lithology intelligent identification method based on a generative adversarial network provided by the present invention. DETAILED DESCRIPTION

[0040] The present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0041] like Figure 1 As shown, the present invention provides a lithology intelligent identification method based on a generative adversarial network, which specifically includes the following steps:

[0042] Step 1: Obtain high spatial resolution remote sensing images of the study area and conduct field surveys

[0043] Obtain high spatial resolution remote sensing images of the study area, circle the bare rock locations that can be obtained by visual interpretation on the images, design field verification routes and verification points, conduct field surveys, and verify the lithology information of the outcrop locations. If there are locations that cannot be reached in the designed route, the lithology information can be supplemented by geological data such as geological maps. The actual verification information should be no less than 100 points.

[0044] Step 2: Preprocess the hyperspectral remote sensing data and remove abnormal bands

[0045] The pre-acquired hyperspectral remote sensing data set is preprocessed (radiation calibration, atmospheric correction, image cropping, etc.), the quality of each band image after preprocessing is checked, and the outliers affected by water vapor absorption near 1400, 1900 and 2400nm are removed.

[0046] Step 3: Based on the lithology information combined with field survey and geological data, image samples are delineated on the pre-processed hyperspectral image to obtain the original training set.

[0047] Based on the lithology information combined with field surveys and geological maps, the sample categories are determined, and image samples of each type of lithology are circled on the pre-processed hyperspectral image to obtain lithology images and spectral data. The size of the circled image samples is adjusted to keep the size of the circled image samples consistent to adapt to the network input. The spectral data corresponding to the circled image samples are normalized, and the image and spectral training set obtained is the original training set.

[0048] Step 4: Construct a deep convolutional generative adversarial network (DCGAN) including a generator network G and a discriminator network D

[0049] In order to generate high-quality hyperspectral remote sensing data samples, the present invention adopts a deep convolutional generative adversarial network (DCGAN). The network structure consists of two core parts: a generator network (G) and a discriminator network (D). Figure 2 As shown in the figure, the input of the generator network is random noise data, which is processed through multiple layers of convolution and deconvolution operations to generate "fake" samples close to real data. The discriminant network accepts the generated "fake" samples and actual samples, performs binary classification, and outputs whether the sample is a real sample. During the training process, the generator optimizes its generation ability by minimizing the discrimination error of the discriminator, while the discriminator optimizes its discrimination ability by maximizing its distinction between "fake" samples and "real" samples. Through this adversarial training mechanism, the generator can gradually learn the distribution characteristics of the data and generate high-quality images and sample spectral data samples.

[0050] The basic process of generating adversarial networks is as follows:

[0051] Step 4.1: Input data: Use random noise as the input of the generator; use the actual hyperspectral remote sensing image as the input of the discriminator as the real sample.

[0052] Step 4.2: Generate “fake” samples: The generator generates “fake” sample images through convolutional neural networks (CNN) and multiple layers of convolution and deconvolution operations. The size and spectral characteristics of the “fake” sample images are close to those of real samples.

[0053] Step 4.3: Distinguish true and false samples: The discriminator compares the "false" samples with the real samples, calculates their authenticity scores, outputs whether the sample is a real sample; and optimizes its parameters according to the loss function.

[0054] Step 4.4: Network optimization: The generator and discriminator gradually update their respective parameters during the confrontation process. The generator optimizes its generation ability by minimizing the discrimination error of the discriminator, while the discriminator optimizes its discrimination ability by maximizing its distinction between "fake" samples and "real" samples, so that the generated "fake" samples are getting closer and closer to the real data distribution.

[0055] Step 5: Introduce convolution operations and self-attention mechanisms to optimize the structure of the generative adversarial network

[0056] Based on the initially constructed DCGAN, the present invention further introduces convolution operations and self-attention mechanisms to improve the performance of the generator and discriminator, especially to improve the quality and details of the generated spectral data samples.

[0057] The basic process of optimizing a generative adversarial network is as follows:

[0058] Step 5.1: Generator optimization: The structure of the generator has been improved, and more convolutional layers and batch normalization (BatchNorm) layers have been added on the basis of the original DCGAN. Each layer of convolution operation gradually enlarges the spatial resolution of the feature map through the deconvolution technology, and finally outputs a spectral data image of size 128×128. The deconvolution operation makes the generated image clearer and maintains the consistency of the spectral features by setting the appropriate convolution kernel, step size and padding method.

[0059] Step 5.2: The generator generates "fake" samples: The input of the generator is a uniformly distributed random noise vector. Through layer-by-layer deconvolution operations, a hyperspectral data image with a shape of 128×128 is finally generated. This image is the original feature map (containing the preliminary image features generated by the generator from the noise). Each layer of the generator is batch normalized to ensure the stability of the model training process. At the same time, the generator output layer uses the Tanh activation function to compress the generated data to the range of [-1,1] to meet the standard format of hyperspectral data.

[0060] Step 5.3: Discriminator optimization: The structure of the discriminator is optimized accordingly, adding a transpose convolution layer and introducing a batch normalization layer in the middle layer. The last layer of the discriminator uses the Sigmoid activation function to output a probability value between 0 and 1, indicating the probability that the sample is real data. The other convolutional layers use the Leaky ReLU activation function to help the network better handle negative inputs and prevent the gradient disappearance problem.

[0061] Step 5.4: The discriminator determines true or false: The discriminator accepts the "fake" samples from the generator and the actual hyperspectral samples, performs convolution and pooling operations layer by layer, and outputs the true or false judgment value.

[0062] Step 5.5: Introduce the self-attention mechanism and generate the self-attention feature map: In the network structure, the self-attention mechanism module is introduced to improve the model's ability to capture spatial and spectral features. The self-attention mechanism effectively improves the generator's ability to model complex spectral features by calculating the similarity matrix of the feature map and weightedly fusing feature information from different positions. The specific implementation method is as follows:

[0063] Use two 1×1 convolutions to perform linear transformation on the convolution feature map and convert it into a matrix form;

[0064] By calculating the attention weight of the feature map (i.e., similarity matrix), and obtaining the normalized attention map through the Softmax function;

[0065] Finally, the original feature map in step 5.2 is multiplied and weighted with the attention map to generate the final self-attention feature map.

[0066] This mechanism helps the generator focus more accurately on important spatial and spectral regions, improving the details and spectral consistency of the generated samples.

[0067] Step 6: Generate sample training and quality evaluation to obtain a new generated sample set

[0068] The generator and discriminator are continuously optimized during the training process, and the quality of the generated spectral samples is continuously improved by iteratively updating their parameters. When the generated samples can pass the true judgment of the discriminator and their spectral characteristics are close to the actual samples, it is considered that the generation process has achieved the expected effect, and the generated samples constitute the new generated sample set after training.

[0069] In order to speed up the training process and avoid the model converging to a local optimal solution, the present invention adopts a strategy of multiple generator updates, that is, after each update of the discriminator weight, the generator is updated 5 times. This strategy enables the generator to quickly learn the distribution of spectral data under the feedback of the discriminator, thereby generating more realistic samples.

[0070] Step 7: Build a one-dimensional and two-dimensional convolutional neural network (1D+2DCNN) model based on the joint features of graph and spectrum. The actual sample data in the training set built in the previous step and the newly generated samples form a new data set, which is then divided into a training set and a validation set in a ratio of 7:3. Then mix the new data set images with the original training set images obtained in step 3 as the input of CNN.

[0071] 1D+2D CNN extracts the spatial features of the image by adding a series of alternating 2D convolutional layers and pooling layers, and fuses them with the spectral features extracted by the 1D convolutional layer. Finally, a fully connected layer is used for classification. The structure of the 1D+2D CNN for extracting joint features of graph and spectrum is shown in the figure. Figure 3 As shown in the figure, on the spectral branch, multiple alternating one-dimensional convolutional layers and pooling layers are used to extract spectral features, and finally a fully connected layer is used for classification. The spectral curve of the pixel is input as a one-dimensional vector, and the corresponding one-dimensional convolutional structure is used to extract the spectral information features of the hyperspectral image. The activation function used in the convolution layer is ReLU. On the image branch, the image features are extracted by stacking two-dimensional convolutional layers, activation functions, and pooling layers. The activation function used in the convolution layer is ReLU, and the pooling layer is the commonly used maximum pooling. After extracting the image features and spectral features, the image features are reconstructed into a one-dimensional tensor, and the Concate fusion method is used to connect the image features with the spectral features, and the fully connected layer is used to classify the joint graph-spectrum features. The loss function is calculated using the cross entropy loss function, and the model training uses the SGD optimizer, and the batch size is set to 256.

[0072] After training, the overall accuracy, average accuracy and Kappa coefficient are used to evaluate the accuracy of the model.

[0073] The present invention is described in detail above with reference to the accompanying drawings and embodiments, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge of ordinary technicians in the field without departing from the purpose of the present invention. The contents not described in detail in the present invention can adopt the existing technology.

Claims

1. A lithology intelligent identification method based on generative adversarial network, characterized in that: The method comprises: Step 1: Obtain high spatial resolution remote sensing images of the study area and conduct field surveys; Step 2: Preprocess the hyperspectral remote sensing data and remove abnormal bands; Step 3: Based on the lithology information obtained from field surveys and geological data, image samples are delineated on the pre-processed hyperspectral images to obtain the original training set; Step 4: Construct a deep convolutional generative adversarial network DCGAN including a generator network G and a discriminator network D; Step 5: Introduce convolution operation and self-attention mechanism to optimize the generative adversarial network structure; Step 6: Generate sample training and quality evaluation to obtain a new generated sample set; Step 7: Construct a one-dimensional + two-dimensional convolutional neural network model based on graph-spectrum joint features.

2. The method for intelligent lithology identification based on generative adversarial network according to claim 1 is characterized in that: The step one is specifically as follows: obtaining high spatial resolution remote sensing images of the study area, circling the bare rock locations that can be obtained by visual interpretation on the images, designing field verification routes and verification points, conducting field surveys, verifying the lithology information of the outcrop locations, and if there are inaccessible locations in the designed route, supplementing the lithology information with geological data.

3. The method for intelligent lithology identification based on generative adversarial network according to claim 1 is characterized in that: The hyperspectral remote sensing data preprocessing in step 2 includes: radiation calibration, atmospheric correction, and image clipping; the abnormal bands include 1400nm, 1900nm, and 2400nm.

4. The method for intelligent lithology identification based on generative adversarial network according to claim 1 is characterized in that: The step three is specifically as follows: according to the lithology information combined with the field survey and the geological map, the sample category is determined, and the image sample of each type of lithology is circled on the pre-processed hyperspectral image to obtain the lithology image and spectral data; the size of the circled image sample is adjusted so that the size of the circled image sample remains consistent to adapt to the input of the network; the spectral data corresponding to the circled image sample is normalized, and the obtained image and spectral training set is the original training set.

5. The method for intelligent lithology identification based on generative adversarial network according to claim 1 is characterized in that: The fourth step comprises: Step 4.1: Input data: Use random noise as the input of the generator; use the actual hyperspectral remote sensing image as the input of the discriminator as a real sample; Step 4.2: Generate "fake" samples: The generator generates "fake" sample images through convolutional neural network CNN, after multiple layers of convolution and deconvolution operations. The size and spectral characteristics of the "fake" sample images are close to the real samples; Step 4.3: Distinguish true and false samples: The discriminator compares the "false" sample with the real sample, calculates its authenticity score, and outputs whether the sample is a real sample; Step 4.4: Network optimization: The generator and discriminator gradually update their respective parameters during the confrontation process. The generator optimizes its generation ability by minimizing the discrimination error of the discriminator, while the discriminator optimizes its discrimination ability by maximizing its distinction between "fake" samples and "real" samples, so that the generated "fake" samples are getting closer and closer to the real data distribution.

6. The method for intelligent lithology identification based on generative adversarial network according to claim 1 is characterized in that: The step five comprises: Step 5.1: Generator optimization: Add more convolutional layers and batch normalization layers on the basis of the original DCGAN. Each convolution operation gradually enlarges the spatial resolution of the feature map through the deconvolution technology, and finally outputs a spectral data image of size 128×128; Step 5.2: The generator generates "fake" samples: The input of the generator is a uniformly distributed random noise vector. Through layer-by-layer deconvolution operations, a hyperspectral data image with a shape of 128×128 is finally generated, which is the original feature map. Each layer of the generator is batch normalized to ensure the stability of the model training process. At the same time, the generator output layer uses the Tanh activation function to compress the generated data to the range of [-1,1] to meet the standard format of hyperspectral data. Step 5.3: Discriminator optimization: The structure of the discriminator is optimized accordingly. A transposed convolution layer is added to the original DCGAN, and a batch normalization layer is introduced in the middle layer. The last layer of the discriminator uses the Sigmoid activation function to output a probability value between 0 and 1, indicating the probability that the sample is real data. The other convolutional layers use the Leaky ReLU activation function to help the network better handle negative inputs and prevent the gradient vanishing problem. Step 5.4: The discriminator determines true or false: The discriminator receives the "false" samples from the generator and the actual hyperspectral samples, performs convolution and pooling operations layer by layer, and outputs the true or false judgment value; Step 5.5: Introduce the self-attention mechanism and generate the self-attention feature map: Introduce the self-attention mechanism module in the network structure; use two 1×1 convolutions to perform linear transformation on the convolution feature map and convert it into a matrix form; calculate the attention weight of the feature map, that is, the similarity matrix, and obtain the normalized attention map through the Softmax function; multiply the original feature map in step 5.2 with the attention map and weight them to generate the final self-attention feature map.

7. The method for intelligent lithology identification based on generative adversarial network according to claim 1 is characterized in that: The step six is ​​specifically as follows: the generator and the discriminator are continuously optimized during the training process, and their respective parameters are iteratively updated to continuously improve the quality of the generated spectral samples; when the generated samples can pass the true judgment of the discriminator and their spectral characteristics are close to the actual samples, it is considered that the generation process has achieved the expected effect, and the generated samples constitute the new generated sample set after training.

8. The method for intelligent lithology identification based on generative adversarial network according to claim 7 is characterized in that: The training of the generated samples in step six is ​​specifically as follows: adopting a strategy of multiple updates of the generator, after each update of the discriminator weight, the generator is updated 5 times, so that the generator can quickly learn the distribution of the spectral data under the feedback of the discriminator and generate more realistic samples.

9. The method for intelligent lithology identification based on generative adversarial network according to claim 1, characterized in that: The step 7 is specifically as follows: the actual sample data in step 6 and the newly generated samples are combined into a new data set, and then the new data set image is mixed with the original training set image obtained in step 3 as the input hyperspectral data of CNN; the spectral branch in the hyperspectral data is extracted by using multiple alternating one-dimensional convolutional layers and one-dimensional pooling layers; In the image branch, multiple alternating two-dimensional convolutional layers and two-dimensional pooling layers are used to extract image features and reconstruct the image features into a one-dimensional tensor. The extracted spectral features and image features are fused and connected, and the joint graph-spectrum features are classified using a fully connected layer to construct a one-dimensional + two-dimensional convolutional neural network model based on the joint graph-spectrum features.

10. The method for intelligent lithology identification based on generative adversarial network according to claim 9, characterized in that: The activation function used by the one-dimensional convolution layer and the two-dimensional convolution layer is ReLU, and the fusion connection of the extracted spectral features and image features adopts the Concate fusion method.

Images