Hyperspectral remote sensing data expansion method based on adversarial neural network

Abstract

The invention relates to a hyperspectral remote sensing data expansion method based on an adversarial neural network, and the method comprises the steps: (1), obtaining a same land class spectrum pixel set according to a plurality of land classes, and carrying out the elimination and spectrum normalization processing of abnormal pixels; (2) inputting the fitting spectrum information and the real land type spectrum information into a spectrum adversarial generation network discriminator, outputting a distance between a fitting spectrum and a real spectrum sample, performing back propagation to update the network discriminator, then calculating a root-mean-square error between the fitting spectrum and the real spectrum sample, performing back propagation to update the network generator, and finally obtaining the real land type spectrum information. Iterating to obtain an optimized multi-land-class spectrum generative adversarial network set; and (3) outputting predicted spectrum information by land classes according to the spectrum generative adversarial networks of different land classes, preferably selecting predicted spectrums meeting precision requirements, and constructing hyperspectral pixel libraries of different land classes. According to the invention, the spectral super-resolution optimization adversarial neural network based on land types is designed, and sample support can be provided for remote sensing image spectral super-resolution related research.

Description

Technical Field

[0001] The present invention relates to the research field of remote sensing image spectral depth super-resolution, and in particular to a hyperspectral remote sensing data expansion method based on adversarial neural network. Background Art

[0002] As a kind of high-dimensional geographic information, hyperspectral remote sensing images contain rich geographic spatial and spectral information, characterize the distribution of surface objects, and have extremely important research and application value. With the continuous development of remote sensing technology, this kind of hyperspectral remote sensing image with integrated map and spectrum has been widely used in the fields of fine classification, mineral detection, species identification, etc. However, due to the limitation of the scanning angle of hyperspectral sensors, the image width is small and it is difficult to meet the needs of large-scale remote sensing applications.

[0003] To this end, domestic and foreign scholars have proposed a multispectral and hyperspectral remote sensing image fusion theory to address the problem of hyperspectral remote sensing image width. That is, using multispectral remote sensing images and hyperspectral remote sensing image pairs with overlapping areas, a spectral mapping method between images is established to map the spectral information of the original hyperspectral remote sensing image to the multispectral image, generating a synthetic hyperspectral remote sensing image with large width and high spatial resolution, such as "Wide-width hyperspectral image reconstruction method based on multi / hyperspectral image fusion" (application number 201310256179.2). At the same time, with the continuous development of deep learning technology, it has also achieved good fitting results in the study of remote sensing image spectral super-resolution.

[0004] The quality and quantity of samples in spectral deep super-resolution research are greatly related to the effect of remote sensing image deep fusion method. How to ensure spectral and spatial characteristics while optimizing the training sample set to the greatest extent has become an important research issue in the current remote sensing image spectral super-resolution. At the same time, the existing remote sensing image spectral deep super-resolution sample optimization methods are difficult to effectively solve this problem. For example, the remote sensing image storage methods disclosed in the patents named "Optimization method for making remote sensing image deep learning samples based on GIS spatial data" (application number CN202110421660.7) and "A remote sensing image cyclic classification method based on automatic optimization of training samples" (application number CN201811503807.1) cannot solve the above problems. Therefore, a method is needed to effectively solve the quality and quantity problems of samples in spectral deep super-resolution research. Summary of the invention

[0005] In order to effectively solve the special needs of remote sensing image spectral deep super-resolution for a large number of continuous samples, the patent of this invention proposes an adversarial optimization method for multi-land class spectral libraries of remote sensing images for spectral super-resolution. It aims to meet the large sample needs of remote sensing image spectral deep super-resolution, and designs a spectral super-resolution optimization adversarial neural network based on land class, which can provide sample support for related research on remote sensing image spectral super-resolution.

[0006] The specific content of the present invention is as follows: A method for expanding hyperspectral remote sensing data based on adversarial neural networks comprises the following steps: Step (1), spectrum optimization of hyperspectral remote sensing images based on classification subgraphs: construct classification subgraphs according to different types of objects in hyperspectral remote sensing images, read spectral pixel sets of different objects in the classification subgraphs of hyperspectral images, traverse the spectral pixel sets, remove and normalize the abnormal pixels in each spectral pixel set, extract the main pixels in the set as the sample set of the adversarial network, and use the sample set selected by the spectrum optimization of hyperspectral remote sensing images based on the classification subgraphs as the input sample set of the adversarial network; Step (2), construction and training of spectral adversarial generative network model: for hyperspectral samples selected from different land types, a spectral adversarial generative network model is constructed, the fitted spectrum information and the real land type spectrum information are input into the spectral adversarial generative network discriminator, the discriminator outputs the distribution distance parameter information of the fitted spectrum and the real spectrum, the spectral adversarial generative network discriminator is updated through back propagation, the root mean square error between the fitted spectrum and the real spectrum sample is calculated, the spectral adversarial generative network generator is updated through back propagation, and the optimized multi-land type spectral adversarial generative network set is obtained through iterative optimization; Step (3), generation of land class spectrum library based on spectral distance optimization: Generate adversarial networks for different land class spectra, output predicted spectral information for each land class, calculate the distance between the predicted land class spectrum and the pure pixel spectrum, set the minimum spectral distance threshold, select the predicted spectrum that meets the accuracy requirements, and construct a hyperspectral pixel library for different land classes.

[0007] Furthermore, step (1) includes: Step (11), traversing the pixels of the hyperspectral image in the form of an array, creating a new file for storing the pixels, and storing the pixels of the hyperspectral image in the corresponding memory block according to the land class code in the classification array; Step (12), the stored file is constructed into a hyperspectral image classification sub-map, and the classification sub-map is read to form a spectral pixel set of different ground objects; Step (13), traversing the spectral pixel set, extracting pixels with abnormal reflectance and abnormal spectral distribution in each spectral pixel set, and performing spectral normalization operations on these abnormal pixels; Step (14), based on step (13), uses the minimum distance discrimination method to traverse the processed spectral pixel set, evaluates the spectral similarity between pixels of the same land type, and selects pixels with concentrated spectral domain distribution as the training set of the adversarial neural network by setting the minimum distance threshold.

[0008] Furthermore, the process of removing abnormal pixels and performing spectral normalization operations in step (13) includes: traversing the spectral pixel set, determining whether the spectral reflectance of the pixel is abnormal, and removing pixels with band reflectance values ​​less than 0 and pixels that obviously deviate from the spectral distribution of the main class; traversing and searching for the maximum and minimum reflectance values ​​for the spectra of pixels of different land classes, and if the maximum reflectance value exceeds 100%, performing normalization operations on all pixels of the land class.

[0009] Furthermore, the following formula is used for normalization: ,in, is the normalized data, is the original data, is the maximum value of the band reflectance of the pixel, It is the minimum reflectance value of the pixel band.

[0010] Furthermore, step (2) includes: Step (21), constructing a spectral adversarial generative network model: the spectral adversarial generative network model includes a spectral super-resolution generator network and a spectral similarity discriminator network, the output dimension of the last linear layer of the generator is consistent with the spectral dimension of the pixel, and the generator network and the discriminator network are symmetrical fully connected layer structures; Step (22), inputting the non-negative random noise low-dimensional data into the spectral adversarial generative network generator to generate fitted real ground-type spectral information; Step (23), for different land types, respectively input the generated fitted real spectrum information and the preferred spectrum data of hyperspectral remote sensing images of different land types into the spectral adversarial generative network discriminator set, the discriminator outputs the distance parameter information between the fitted spectrum and the real spectrum, and back-propagates to update the spectral adversarial generative network discriminator; Step (24), calculating the root mean square error between the fitted spectrum and the true spectrum sample, outputting the error information between the two, back-propagating to update the spectrum adversarial generative network generator, and iteratively optimizing and updating the multi-ground spectrum adversarial generative network set; Step (25), for each land class spectrum adversarial generative network, set the spectral minimum distance threshold, calculate the spectral angular distance between the predicted pixel set generated by the training spectral super-resolution generator network during each iterative optimization and the training pixel set selected in step (14), and save the model parameters of the land class when the spectral minimum threshold requirement is met.

[0011] Furthermore, in step (23), the spectral adversarial generative network generator is fixed, and the network discriminator is updated iteratively for multiple times before calculating the root mean square error between the fitted spectrum and the true spectrum sample, and the root mean square error information between the two is output for back-propagation to update the spectral adversarial generative network generator.

[0012] Furthermore, step (3) includes: Step (31), using the different land type pixel spectra adversarial generative network generator constructed in step 2 to predict sufficient land type similar spectra respectively; Step (32), using the spectral similarity index, setting the pixel spectrum minimum distance threshold, selecting the predicted spectral pixel that meets the accuracy requirement as the generated spectrum of the current land class, and repeating steps (31)-(32) to obtain the predicted spectral data of all land classes on the image; Step (33), based on the land class sub-map concept, according to the given image size, the spectral sample library cubes of different land classes are spliced.

[0013] The multi-terrestrial spectral library adversarial optimization method for remote sensing images for spectral super-resolution of the present invention meets the pixel requirements of spectral depth super-resolution learning of multi-spectral and hyperspectral remote sensing images, and can greatly improve the provision of high-quality, large-scale and continuous sample support required for spectral depth super-resolution of remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The specific implementation of the present invention will be further explained below in conjunction with the accompanying drawings.

[0015] Figure 1 is a flow chart of the present invention; Figure 2 To invent a spectral optimization process diagram for hyperspectral remote sensing images based on classification sub-graphs.

[0016] Figure 3 This is a network structure diagram of the spectral adversarial generation model of the present invention.

[0017] Figure 4 A process diagram for generating a total land-based spectral library of the present invention; Figure 5a It is the Indian Pines classification subgraph in the embodiment of the present invention; Figure 5b The hyperspectral pixel library expansion results of different land types are constructed by optimizing samples from classified sub-maps for the Indian Pines dataset. DETAILED DESCRIPTION

[0018] Combination Figure 1-Figure 4 The present invention provides a method for adversarial optimization of multi-land-class spectral libraries of remote sensing images for spectral super-resolution, including spectral optimization of hyperspectral remote sensing images based on classified subgraphs, construction and training of spectral adversarial generation network models, and generation of land-class spectral libraries based on spectral distance optimization, as follows: Step (1), spectral optimization of hyperspectral remote sensing images based on classification subgraphs: construct classification subgraphs according to different types of objects in hyperspectral remote sensing images, read spectral pixel sets of different objects in the classification subgraphs of hyperspectral images, traverse the spectral pixel sets, remove and normalize the pixels with abnormal reflectance and spectral distribution in each spectral pixel set, and extract the main pixels in the set as the sample set of the adversarial network, and use the sample set selected based on the classification subgraph spectral optimization of hyperspectral remote sensing images as the input sample set of the adversarial network; Step (2), spectral adversarial generative network model construction and training: for different land-type selected hyperspectral samples, a spectral adversarial generative network model is constructed. Non-negative random noise data is used to fit the spectral information of the land-type to be predicted, and the fitted spectral information and the real land-type spectral information are input into the spectral adversarial generative network discriminator. The discriminator outputs the distribution distance parameter information of the fitted spectrum and the real spectrum, and the spectral adversarial generative network discriminator is updated by back propagation. The root mean square error between the fitted spectrum and the real spectrum sample is then calculated, and the spectral adversarial generative network generator is updated by back propagation. The optimized multi-land-type spectral adversarial generative network set is obtained by iterative optimization.

[0019] Step (3), generation of land class spectrum library based on spectral distance optimization: for different land class spectra, an adversarial generative network is used to output predicted spectral information for each land class by inputting non-negative random noise into the adversarial neural network, calculating the distance between the predicted land class spectrum and the pure pixel spectrum, setting the minimum spectral distance threshold, optimizing the predicted spectrum that meets the accuracy requirements, and constructing a hyperspectral pixel library for different land classes.

[0020] In this embodiment, the specific process of step (1) includes: Step (11), traversing the pixels of the hyperspectral image in the form of an array, and creating a new file to store the pixels, storing the pixels of the hyperspectral image into corresponding memory blocks according to the land class codes in the classification array, and creating a new file to store the pixels; Step (12), the stored file is constructed into a classified hyperspectral image classification sub-map, and the classification sub-map is read to form a set of spectral pixels of different ground objects; Step (13), traverse the spectral pixel set, extract the pixels with abnormal reflectance and abnormal spectral distribution in each spectral pixel set, and remove and perform spectral normalization operations on these abnormal pixels.

[0021] Step (14), based on step (13), uses the minimum distance discrimination method to traverse the processed spectral pixel set, evaluates the spectral similarity between pixels of the same land type, and selects pixels with concentrated spectral domain distribution as the training set of the adversarial neural network by setting the minimum distance threshold.

[0022] The specific process of removing abnormal pixels and normalizing the spectrum in step (13) is as follows: traverse the spectral pixel set to determine whether the pixel spectral reflectance is abnormal. Pixels with band reflectance values ​​less than 0 and pixels that obviously deviate from the main class spectral distribution are removed, and pixels with band reflectance greater than 1 are normalized.

[0023] In this embodiment, the specific process of step (2) of constructing and training the spectral adversarial generative network model includes: Step (21), the model consists of a spectral super-resolution generator network and a spectral similarity discriminator network. Both the spectral super-resolution generator network and the spectral similarity discriminator network are composed of a fully connected network and an activation function ELU. The generator network and the discriminator network must be a symmetrical fully connected layer structure, and the last layer of the generator is composed of an activation function Tanh. Step (22), inputting non-negative random noise low-dimensional data into the spectral adversarial generative network generator to generate fitted real land class spectral information, for different land classes, respectively inputting the generated fitted real spectral information and the preferred different land class hyperspectral remote sensing image spectral data into the spectral adversarial generative network discriminator set, the discriminator outputs the distance parameter information between the fitted spectrum and the real spectrum, and back-propagating to update the spectral adversarial generative network discriminator; Step (23), calculate the root mean square error between the fitted spectrum and the true spectrum sample and output the error information between the two, back-propagate and update the spectrum adversarial generative network generator, and iteratively optimize and update the multi-ground spectrum adversarial generative network set.

[0024] Step (24), for each land class spectrum adversarial generative network, set a spectral minimum distance threshold, calculate the spectral angular distance between the predicted pixel set generated by the training spectral super-resolution generator network during each iterative optimization and the training pixel set selected in step (14), and save the model parameters of the land class when the spectral minimum threshold requirement is met.

[0025] In this embodiment, preferably, in step (23), the spectral adversarial generative network generator is fixed, and after updating the network discriminator for multiple iterations, the root mean square error between the fitted spectrum and the true spectrum sample is calculated and the root mean square error information between the two is output to perform back propagation to update the spectral adversarial generative network generator.

[0026] In this embodiment, the specific process of step (3) of selecting the predicted spectrum that meets the accuracy requirement and constructing the hyperspectral pixel library of different land types includes: Step (31), using the different ground type pixel spectrum adversarial expansion model constructed in step 2, respectively predicting sufficient similar spectra of common ground types; Step (32), using the spectral similarity index, setting the pixel spectrum minimum distance threshold, selecting the predicted spectral pixel that meets the accuracy requirement as the generated spectrum of the current land class, and repeating steps (31)-(32) to obtain the predicted spectral data of all land classes on the image; Step (33), based on the idea of ​​classified sub-graphs, according to the given image size, the spectral sample library cubes of different land types are spliced.

[0027] This example uses Windows 10 as the operating system environment, PyCharm 2022.2.1 x64 as the development environment, Python as the development language, and the hyperspectral file is stored as a TIF file. The Indian Pines dataset and TIF file are used as sample data. The specific implementation process is as follows: 1. Spectral optimization of hyperspectral remote sensing images based on classified subgraphs: Figure 2 As shown, the specific steps are: Step 1: Construct a hyperspectral classification subgraph Input the hyperspectral image and the classified image (In this example, the Indian Pines dataset and the classified image are both tif files), use gdal.ReadAsArray to read the hyperspectral image and the classified image and convert them into arrays and store them in memory, and use the gdal.GetDriverByName method to create a new TIF file with the same number of bands, projections, and other information as the original hyperspectral image, with the number of rows and columns being 1×1. Use the WriteArray function to read the pixel values ​​of each band of the traversed hyperspectral image pixel into the TIF file, and name the TIF file as the coordinate position of the pixel. For the hyperspectral pixel vector set under different land types, calculate the maximum square size that can be formed, and construct the hyperspectral land type submap respectively.

[0028] Step 2: Construct spectral pixel sets of different objects Create folders for the corresponding features according to different feature categories, and then use gdal.ReadAsArray to read the hyperspectral land classification submap. Traverse the classification submap, and use the gdal.GetDriverByName method for each element in the traversed array to create a new TIF file with the same number of bands, projections, and other information as the original hyperspectral image, with 1×1 rows and columns. According to the feature codes of the classification submap, they are all read into the TIF file and placed in different folders for the corresponding features.

[0029] Step 3: Optimize the hyperspectral pixel set Read the folders containing TIF files of different objects, use the ReadAsArray method to read these TIF files and convert them into arrays and store them in memory, read the pixel values ​​of each band in the array, and determine whether the band reflectance value is less than 0 or greater than 1. If it is less than 0, it will be directly eliminated, and if it is greater than 1, the formula is used ,in, is the normalized data, is the original data, is the maximum value of the band reflectance of the pixel, is the minimum reflectance of the pixel band. Perform spectral normalization and then overwrite the original TIF file with the normalized array.

[0030] Step 4: Select pixels with concentrated spectral distribution Use SAM (spectral angle) in the imgvision library as the evaluation index, set the minimum distance threshold to select pixels with concentrated spectral domain distribution, traverse each feature folder, traverse each TIF file in the file, use the plt method in the python library to display the distribution of spectral curves of the corresponding feature folder, select typical pixel curves of the feature, read each TIF file as an array and use the SAM method with the typical pixel array of the feature, save TIF files with a distance less than the minimum distance threshold, and remove TIF files with a distance greater than the minimum distance threshold. The folders storing TIF files of multiple land types obtained at the end are used as the preferred spectral pixel set.

[0031] 2. Spectral adversarial generative network model construction and training, such as Figure 3 As shown, it includes the following steps: Step 1: Build a spectral adversarial generative network model The model consists of a spectral super-resolution generator network and a spectral similarity discriminator network. In order to ensure that the output of the designed spectral adversarial generation network is consistent with the spectral dimension of the pixel, the output dimension of the last linear layer of the generator is designed to be consistent with the spectral dimension of the pixel. At the same time, the linear layers of the generator and discriminator of the spectral adversarial generation network are designed to be symmetrical. The activation function in the middle of each linear layer is replaced with the ELU activation function that can better converge the network. The last layer of the generator network uses the Than activation function to speed up convergence and reduce the occurrence of gradient diffusion and gradient vanishing.

[0032] Step 2: Train the spectral adversarial network discriminator Set the spectral adversarial network training hyperparameters (number of trainings, number of fixed discriminators, learning rate, gradient clipping size, size of training data channel dimensions, batch size for each training), initialize the Adam optimizer, and before starting training, traverse the folders storing TIFs of different land types to form a dataloader object array for iterative training based on the batch size. Before reaching the maximum number of trainings, first fix the spectral adversarial network generator, use the numpy.random.random method in the numpy library to generate a non-negative random noise array to input the spectral adversarial network generator, input the fitted spectral information array and the real spectral information array read from the dataloader into the spectral adversarial network discriminator, and then the discriminator outputs the distance parameter information between the fitted spectrum and the real spectrum distribution, and back-propagates to update the spectral adversarial network discriminator.

[0033] Step 3: Train the spectral adversarial network generator When the set number of training discriminators is reached, the root mean square error between the fitted spectral information array and the real spectral information array read from the dataloader is calculated, and the spectral adversarial generative network generator is updated by backpropagation. The training is iterated continuously until the set maximum number of iterations is reached.

[0034] Step 4: Save model parameters for different land types A spectral minimum distance threshold is set for the spectrum adversarial generative network of each land type. The spectral angular distance is calculated between the predicted pixel set generated by the training spectral super-resolution generator network during each iterative optimization and the training pixel set selected from the hyperspectral remote sensing image spectrum optimization based on the classification sub-graph. When the spectral minimum threshold requirement is met, the torch.save method is used to save the model parameter .pth file of the land type.

[0035] 3. Generation of land-type spectral library based on spectral distance optimization, such as Figure 4 As shown, it includes the following steps: Step 1: Predict sufficient similar spectral pixel libraries for different land types Create different land feature expansion result folders according to different land types, use the trained spectral adversarial network generator to load the .pth files of different land types, input the non-negative random noise array of the specified dimension to the generator one by one, generate the predicted spectral array, use the gdal.GetDriverByName method, create a TIF file with the same number of bands, projection and other information as the original hyperspectral image, and the number of rows and columns is 1×1, and save it to the land feature expansion result folder of the specified land type.

[0036] Step 2: Select the generated spectral pixels that meet the accuracy requirements Use SAM (spectral angle) in the imgvision library as the evaluation index, set the minimum distance threshold of the pixel spectrum, traverse all the TIF files in the expansion result folder, and use the pixel spectra less than the threshold as the generated spectral pixels of the current land class. Pixel spectra greater than the threshold are removed, and repeat steps 1 to 2 to obtain TIF folders of predicted spectral pixels of different land classes.

[0037] Step 3: Splice to form spectral sample library cubes of different land types Based on the maximum square size that the original classification sub-map can form, the idea of ​​forming a hyperspectral land class sub-map is constructed separately. Use the gdal.GetDriverByName method to create a large-map TIF file with the same number of bands, projections and other information as the original hyperspectral image, and the same number of rows and columns as the original land class sub-map. Traverse and read all the folders storing the expanded TIF files, set the maximum square size of the feature in the large-map TIF file according to the read feature folder, and then use the gdal.ReadAsArray method to write the TIF file into the set maximum square. After traversal, a hyperspectral TIF file containing a large number of continuous, high-quality features and clear blocks is formed.

[0038] like Figure 5a , 5b The figure shows a schematic diagram of constructing a hyperspectral pixel library of different land types by selecting samples from the classification sub-graphs for the Indian Pines dataset of this embodiment, where: Figure 5a is the Indian Pines classification subgraph, Figure 5b The results of the hyperspectral pixel library expansion for different land types. Figure (a) is a whole image with obvious boundary lines, which is used to clearly compare the differences between different land objects in the classification sub-image. Figure (b) is the result of the hyperspectral pixel library expansion for different land types based on the classification sub-image.

[0039] It can be seen that the hyperspectral remote sensing data expansion method based on adversarial neural network in this application can better meet the needs of continuous, regular, and high-quality samples in deep learning in remote sensing image applications. The method of the present invention is aimed at the pixel requirements of spectral deep super-resolution learning of multi-spectral and hyperspectral remote sensing images, and solves the problem that it is difficult to obtain a set of pixels that are spatially continuous and of the same type in a single hyperspectral remote sensing image. It can greatly improve the quality and quantity of samples required for spectral deep super-resolution of remote sensing images, and provide high-quality, large-scale, and continuous sample support for research in the field of spectral super-resolution of remote sensing images.

[0040] Many specific details are described in the above description to facilitate a full understanding of the present invention. However, the above description is only a preferred embodiment of the present invention. The present invention can be implemented in many other ways different from those described herein, so the present invention is not limited to the specific implementation disclosed above. At the same time, any person familiar with the art can make many possible changes and modifications to the technical solution of the present invention using the methods and technical contents disclosed above without departing from the scope of the technical solution of the present invention, or modify it into an equivalent embodiment of equivalent changes. Any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of protection of the technical solution of the present invention.

Claims

1. A hyperspectral remote sensing data expansion method based on adversarial neural network, characterized by: The following steps are involved: Step (1), spectrum optimization of hyperspectral remote sensing images based on classification subgraphs: construct classification subgraphs according to different types of objects in hyperspectral remote sensing images, read spectral pixel sets of different objects in the classification subgraphs of hyperspectral images, traverse the spectral pixel sets, remove and normalize the abnormal pixels in each spectral pixel set, extract the main pixels in the set as the sample set of the adversarial network, and use the sample set selected by the spectrum optimization of hyperspectral remote sensing images based on the classification subgraphs as the input sample set of the adversarial network; Step (2), construction and training of spectral adversarial generative network model: for hyperspectral samples selected from different land types, a spectral adversarial generative network model is constructed, the fitted spectrum information and the real land type spectrum information are input into the spectral adversarial generative network discriminator, the discriminator outputs the distribution distance parameter information of the fitted spectrum and the real spectrum, the spectral adversarial generative network discriminator is updated through back propagation, the root mean square error between the fitted spectrum and the real spectrum sample is calculated, the spectral adversarial generative network generator is updated through back propagation, and the optimized multi-land type spectral adversarial generative network set is obtained through iterative optimization; Step (3), generation of land class spectrum library based on spectral distance optimization: Generate adversarial networks for different land class spectra, output predicted spectral information for each land class, calculate the distance between the predicted land class spectrum and the pure pixel spectrum, set the minimum spectral distance threshold, select the predicted spectrum that meets the accuracy requirements, and construct a hyperspectral pixel library for different land classes.

2. The method for hyperspectral remote sensing data expansion based on adversarial neural network according to claim 1 is characterized in that: Step (1) includes: Step (11), traversing the pixels of the hyperspectral image in the form of an array, creating a new file for storing the pixels, and storing the pixels of the hyperspectral image in the corresponding memory block according to the land class code in the classification array; Step (12), the stored file is constructed into a hyperspectral image classification sub-map, and the classification sub-map is read to form a spectral pixel set of different ground objects; Step (13), traversing the spectral pixel set, extracting pixels with abnormal reflectance and abnormal spectral distribution in each spectral pixel set, and performing spectral normalization operations on these abnormal pixels; Step (14), based on step (13), uses the minimum distance discrimination method to traverse the processed spectral pixel set, evaluates the spectral similarity between pixels of the same land type, and selects pixels with concentrated spectral domain distribution as the training set of the adversarial neural network by setting the minimum distance threshold.

3. The method for hyperspectral remote sensing data expansion based on adversarial neural network according to claim 2 is characterized in that: The process of removing abnormal pixels and performing spectral normalization operations in step (13) includes: traversing the spectral pixel set, determining whether the pixel spectral reflectance is abnormal, and removing pixels with band reflectance values ​​less than 0 and pixels that obviously deviate from the main class spectral distribution; traversing and searching for the maximum and minimum reflectance values ​​for the spectra of pixels of different land classes, and if the maximum reflectance value exceeds 100%, performing normalization operations on all pixels of the land class.

4. The method for hyperspectral remote sensing data expansion based on adversarial neural network according to claim 3 is characterized in that: The following formula is used for normalization: ,in, is the normalized data, is the original data, is the maximum value of the band reflectance of the pixel, It is the minimum reflectance value of the pixel band.

5. The method for hyperspectral remote sensing data expansion based on adversarial neural network according to claim 2 is characterized in that: Step (2) includes: Step (21), constructing a spectral adversarial generative network model: the spectral adversarial generative network model includes a spectral super-resolution generator network and a spectral similarity discriminator network, the output dimension of the last linear layer of the generator is consistent with the spectral dimension of the pixel, and the generator network and the discriminator network are symmetrical fully connected layer structures; Step (22), inputting the non-negative random noise low-dimensional data into the spectral adversarial generative network generator to generate fitted real ground-type spectral information; Step (23), for different land types, respectively input the generated fitted real spectrum information and the preferred spectrum data of hyperspectral remote sensing images of different land types into the spectral adversarial generative network discriminator set, the discriminator outputs the distance parameter information between the fitted spectrum and the real spectrum, and back-propagates to update the spectral adversarial generative network discriminator; Step (24), calculating the root mean square error between the fitted spectrum and the true spectrum sample, outputting the error information between the two, back-propagating to update the spectrum adversarial generative network generator, and iteratively optimizing and updating the multi-ground spectrum adversarial generative network set; Step (25), for each land class spectrum adversarial generative network, set the spectral minimum distance threshold, calculate the spectral angular distance between the predicted pixel set generated by the training spectral super-resolution generator network during each iterative optimization and the training pixel set selected in step (14), and save the model parameters of the land class when the spectral minimum threshold requirement is met.

6. The method for hyperspectral remote sensing data expansion based on adversarial neural network according to claim 5 is characterized in that: In step (23), the spectral adversarial generative network generator is fixed, and after updating the network discriminator through multiple iterations, the root mean square error between the fitted spectrum and the true spectrum sample is calculated, and the root mean square error information between the two is output for back propagation to update the spectral adversarial generative network generator.

7. The method for hyperspectral remote sensing data expansion based on adversarial neural network according to claim 5 is characterized in that: Step (3) includes: Step (31), using the different land type pixel spectra adversarial generative network generator constructed in step 2 to predict sufficient land type similar spectra respectively; Step (32), using the spectral similarity index, setting the pixel spectrum minimum distance threshold, selecting the predicted spectral pixel that meets the accuracy requirement as the generated spectrum of the current land class, and repeating steps (31)-(32) to obtain the predicted spectral data of all land classes on the image; Step (33), based on the land class sub-map concept, according to the given image size, the spectral sample library cubes of different land classes are spliced.

Images