A compressed sensing hyperspectral data reconstruction method based on deep learning

Abstract

A method for reconstructing compressed sensing hyperspectral data based on deep learning, characterized in that it includes the following steps: Step 1: training a compressed sensing hyperspectral data reconstruction network model based on a convolutional neural network; Step 2: reconstructing the compressed sensing hyperspectral data using the reconstruction network model trained in Step 1. The present invention uses a convolutional neural network to solve the problem of reconstructing compressed sensing hyperspectral data, and learns the inverse transformation from compressed measurement data to original data by training the network, and further uses the trained model to reconstruct the compressed sensing hyperspectral data, thereby achieving fast and accurate reconstruction of compressed sensing hyperspectral data. Compared with traditional iterative reconstruction algorithms, the reconstruction quality is improved, and the calculation time for reconstructing hyperspectral data is significantly improved, which is much faster than traditional algorithms.

Description

Technical Field

[0001] The present invention belongs to the technical field of hyperspectral image compressed sensing information reconstruction technology processing, and specifically relates to a compressed sensing hyperspectral data reconstruction method based on deep learning. Background Art

[0002] Hyperspectral data can obtain spatial and spectral information at the same time, providing richer information for target identification, and is widely used in remote sensing monitoring, biomedical imaging, precision agriculture, environmental monitoring and other fields. However, the high dimensionality of hyperspectral data brings great challenges to its storage and transmission.

[0003] Compressed sensing (CS) can collect signals at a sampling rate far lower than the Nyquist rate and reconstruct the original signal through a reconstruction algorithm, thus saving a lot of resources. In recent years, compressed sensing spectral imaging technology has received widespread attention. Based on the theory of compressed sensing, the complete spatial information and spectral information can be restored through less sampling measurement data, which can effectively solve the problem of hyperspectral data storage and transmission. The current compressed sensing spectral imaging technology mainly uses an iterative optimization algorithm in the information reconstruction stage, which has the problems of long reconstruction time and low reconstruction accuracy. Summary of the invention

[0004] In order to solve the above-mentioned technical problems, the present invention designs a compressed sensing hyperspectral data reconstruction method based on deep learning.

[0005] In order to solve the above-mentioned technical problems, the present invention adopts the following solutions:

[0006] A method for reconstructing compressed sensing hyperspectral data based on deep learning, comprising the following steps:

[0007] Step 1: Train the compressed sensing hyperspectral data reconstruction network model based on convolutional neural network;

[0008] Step 2: Use the reconstruction network model trained in step 1 to reconstruct the compressed sensing hyperspectral data.

[0009] Further, the step 1 comprises:

[0010] Step 11: Design the measurement matrix Φ;

[0011] Step 12: Select an original hyperspectral data set, select K original hyperspectral data as the training set, and select S original hyperspectral data as the validation set; use the measurement matrix Φ to compress and sample the original hyperspectral data x in the training set and the validation set to obtain compressed data y;

[0012] Step 13: Divide the compressed data y and the original hyperspectral data x into blocks; Divide into a set of small cubes x1, x2, ..., x i ,in L x ×L y and N λ are the spatial resolution and number of bands of the original hyperspectral data, N x ×N y is the spatial resolution of the cube; similarly, the data will be compressed Divide into y1,y2,…,y i ,in M λ is the number of bands of compressed data; compressed data block y1,y2,…,y i With the original hyperspectral data block x1,x2,…,x i One-to-one correspondence forms training data pairs, making each size N x ×N y ×M λ Each compressed data block has its corresponding true value N x ×N y ×N λ ;

[0013] Step 14: Construct a compressed sensing hyperspectral data reconstruction network model;

[0014] Step 15: Input the training data pairs in step 13 into the reconstruction network model in step 14, and compress the data blocks y1, y2, …, y i By reconstructing each convolutional layer of the network model, the compressed data blocks y1, y2, ..., y i Recovering from low-dimensional measurement data to high-dimensional reconstructed data

[0015] Step 16: Calculate the error between the reconstructed data and the original hyperspectral data through the loss function and perform back propagation to update the parameters of the reconstruction network model in step 14;

[0016] Step 17: Repeat steps 15-16 until the error calculated by the loss function reaches a satisfactory value, and then end the training. At this time, the trained reconstructed network model is obtained.

[0017] Furthermore, in step 11, the measurement matrix Φ adopts a random Gaussian measurement matrix or a Bernoulli matrix.

[0018] Furthermore, in step 12, the hyperspectral data in the original hyperspectral data set is the original hyperspectral data after spatial downsampling processing.

[0019] Furthermore, in step 12, only the spectral dimension of the original hyperspectral data x is compressed, and the spatial resolution remains unchanged. The processing method of the compressed sampling of the original hyperspectral data x is y=Φx, where the measurement matrix is

[0020] Further, the step 14 comprises:

[0021] Step 141: obtaining initial reconstructed data by pseudo-inverse operation; the function of pseudo-inverse operation is to reversely project the hyperspectral data from the compression domain to the spectral domain;

[0022] The pseudo-inverse operation is performed using N λ This is achieved by a 2D convolutional layer with a convolution kernel size of 1×1. This operation generates the initial reconstructed data, and the weight of the convolutional layer is determined by the pseudo-inverse matrix. The calculation process of the pseudo-inverse matrix is:

[0023]

[0024] Setting this layer to non-trainable during network training can avoid overfitting;

[0025] Step 142: The initial reconstructed data is subjected to feature extraction by the Vnet model.

[0026] Further, in step 14, Vnet includes an encoder and a decoder, and a connection is added between the encoder and the decoder:

[0027] The structure of the encoder module is: residual block-downsampling-residual block-downsampling-residual block-downsampling-residual block;

[0028] The first residual block contains 2 convolutional layers, each of which uses a 3×3×7 convolution kernel with a stride of 1, and the number of convolution kernels is 8;

[0029] The second residual block contains 2 convolutional layers, each of which uses a 3×3×5 convolution kernel with a stride of 1, and the number of convolution kernels is 8;

[0030] The third residual block contains 2 convolutional layers, each of which uses a 3×3×5 convolution kernel with a stride of 1, and the number of convolution kernels is 32;

[0031] The fourth residual block contains two convolutional layers, each of which uses a 3×3×3 convolution kernel, and the number of convolution kernels is 128;

[0032] The first, second, and third residual blocks are followed by a maximum pooling layer with a convolution kernel size of 1×1×2;

[0033] The structure of the decoder module is: upsampling-residual block-upsampling-residual block-upsampling-residual block:

[0034] A upsampling layer with a convolution kernel of (1,1,2) is added before each residual block;

[0035] The first residual block contains a concatenation and two convolutional layers. Each convolutional layer uses a 3×3×5 convolution kernel with a stride of 1, and the number of convolution kernels is 32.

[0036] The second residual block contains a concatenation and two convolutional layers. Each convolutional layer uses a 3×3×5 convolution kernel with a stride of 1, and the number of convolution kernels is 16.

[0037] The third residual block contains a concatenation and two convolutional layers. Each convolutional layer uses a 3×3×7 convolution kernel with a stride of 1, and the number of convolution kernels is 8.

[0038] The last convolution layer uses a 1×1×1 convolution kernel with a stride of 1 and a number of convolution kernels of 1;

[0039] Each convolution layer uses 3D convolution. To avoid blocking effects, reflection padding is used before each 3D convolution, and the ReLU activation function is used after each 3D convolution.

[0040] Furthermore, in step 16, the loss function uses a mean square error function MSE to calculate the loss value of each original hyperspectral data and its corresponding reconstructed data;

[0041] The total loss value is calculated and optimized using the ADAM optimizer with a learning rate of 0.001.

[0042] Furthermore, in step 17, the reconstruction result is evaluated using the peak signal-to-noise ratio (PSNR) to assess the accuracy of the reconstructed data:

[0043]

[0044] Among them, MSE represents the mean square error between the original image and the reconstructed image. The larger the PSNR index value, the better the reconstruction effect. n represents the quantization bit depth of the image, and n is generally a positive integer.

[0045] Furthermore, the method also includes step 18: selecting M original hyperspectral data from the original hyperspectral data set in step 12 as a test set, wherein M=S, that is, the test set and the validation set contain the same amount of hyperspectral data; using the measurement matrix Φ to compress and sample the original hyperspectral data in the test set to obtain test compressed data sets; using the block processing method in step 13 to block the data in the test set and the test compressed data set to form test data; inputting the test data into the trained reconstruction network model obtained in step 17 for testing, if the test is qualified, the trained reconstruction network model is put into use, otherwise, steps 11-18 are repeated.

[0046] Further, the step 2 comprises:

[0047] Step 21: compress and sample the hyperspectral data using the measurement matrix Φ to obtain compressed hyperspectral data Y; the measurement matrix Φ adopts the measurement matrix Φ designed in step 1;

[0048] Step 22: Divide the compressed hyperspectral data Y into blocks to obtain Y1, Y2, …, Y i ; The block processing method is the same as the block processing method of step 13 in step 1;

[0049] Step 23: Use the trained reconstruction network model to reconstruct Y1, Y2, …, Y i Reconstruction is performed, and the reconstructed image blocks are rearranged in sequence to obtain a reconstructed image.

[0050] The compressed sensing hyperspectral data reconstruction method based on deep learning has the following beneficial effects:

[0051] (1) In the present invention, the first layer in the reconstruction network uses a pseudo-inverse operation to generate a preliminary initial reconstructed image, which is more conducive to obtaining a reconstruction effect close to the original image.

[0052] (2) In the present invention, a convolutional neural network is used to solve the problem of compressed sensing hyperspectral data reconstruction. The network is trained to learn the inverse transformation of compressed sampled measurement data to original data, and the trained model is further used to reconstruct the compressed hyperspectral data. This achieves fast and accurate reconstruction of compressed sensing hyperspectral data. Compared with traditional iterative reconstruction algorithms, the reconstruction quality is improved, and the calculation time for reconstructing hyperspectral data is significantly improved, which is much faster than traditional algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 : Flow chart of a method for reconstructing compressed sensing hyperspectral data based on deep learning in an embodiment of the present invention;

[0054] Figure 2 : Diagram of the training data set block processing process in an embodiment of the present invention;

[0055] Figure 3 : A network diagram of compressed sensing hyperspectral data reconstruction based on deep learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The present invention will be further described below in conjunction with the accompanying drawings:

[0057] Figures 1 to 3 A specific implementation of a compressed sensing hyperspectral data reconstruction method based on deep learning of the present invention is shown. Figure 1 is a flow chart of a method for reconstructing compressed sensing hyperspectral data based on deep learning in this embodiment; Figure 2 is a diagram of the training data set block processing process in this embodiment; Figure 3 This is a compressed sensing hyperspectral data reconstruction network diagram based on deep learning in this implementation.

[0058] like Figures 1 to 3 As shown, the compressed sensing hyperspectral data reconstruction method based on deep learning in this embodiment includes the following steps:

[0059] Step 1: Train the compressed sensing hyperspectral data reconstruction network model based on convolutional neural network;

[0060] Step 2: Use the reconstruction network model trained in step 1 to reconstruct the compressed sensing hyperspectral data.

[0061] Specifically, step 1 includes the following steps:

[0062] Step 11: Construct the measurement matrix Φ;

[0063] The design purpose of the measurement matrix is ​​to sample the original signal to obtain M λ measurements, and it is guaranteed that a length N can be reconstructed from them λ The original signal x. The present invention adopts a random Gaussian measurement matrix.

[0064] Step 12: Select an original hyperspectral data set, use K original hyperspectral data as the training set, and select S original hyperspectral data as the validation set;

[0065] Because the spatial dimension of the directly collected hyperspectral data is large, it can generate thousands of blocks, and the storage also requires a large amount of memory, which will result in a long training time. Therefore, it is impossible to perform batch processing in the neural network. Therefore, the directly collected hyperspectral data is spatially downsampled to change the spatial resolution of the hyperspectral data, and the processed hyperspectral data is used as the original hyperspectral data.

[0066] The original hyperspectral data x in the training set and validation set are compressed and sampled through the measurement matrix Φ to obtain the measurement data y;

[0067] In the present invention, only the spectral dimension is compressed, the spatial resolution remains unchanged, and the hyperspectral data is sampled pixel by pixel using the measurement matrix. The sampling processing method is as follows:

[0068] y j =Φx j

[0069] is the value of a certain pixel; is the compressed value.

[0070] In this embodiment, the original hyperspectral data size is The size of the compressed hyperspectral data after sampling is

[0071] Step 13: In order to reduce storage and speed up training, block training is adopted. Split into a set of small cubes of the same size x1, x2, ..., x i ,in L x ×L y and N λ are the spatial resolution and number of bands of hyperspectral data, N x ×N y is the spatial resolution of the small cube. Similarly, the compressed data Divide into y1,y2,…,y i ,in M λ is the number of bands of compressed hyperspectral data. Compressed data block y1,y2,…,y i With the original hyperspectral data block x1,x2,…,x i One-to-one correspondence forms training data pairs, making each size N x ×N y ×N λ Each compressed data block has its corresponding true value N x ×N y ×M λ ;

[0072] Each hyperspectral data is divided into blocks with overlaps.

[0073] Step 14: Construct a compressed sensing hyperspectral data reconstruction network model;

[0074] The compressed sensing hyperspectral data reconstruction network model proposed in the present invention includes pseudo-inverse operation and Vnet structure:

[0075] The function of the pseudo-inverse operation is to reversely project the data in the compressed domain into the spectral domain;

[0076] The pseudo-inverse operation is performed using N λ This is achieved by a 2D convolutional layer with a convolution kernel size of 1×1. This operation generates the initial reconstructed data, and the weight of the convolutional layer is determined by the pseudo-inverse matrix. The calculation process of the pseudo-inverse matrix is:

[0077]

[0078] Setting this layer to non-trainable during network training can avoid overfitting;

[0079] Vnet consists of an encoder and a decoder, with a connection between them:

[0080] The structure of the encoder module is: residual block-downsampling-residual block-downsampling-residual block-downsampling-residual block;

[0081] The first residual block contains 2 convolutional layers, each of which uses a 3×3×7 convolution kernel with a stride of 1, and the number of convolution kernels is 8;

[0082] The second residual block contains 2 convolutional layers, each of which uses a 3×3×5 convolution kernel with a stride of 1 and 16 convolution kernels;

[0083] The third residual block contains 2 convolutional layers, each of which uses a 3×3×5 convolution kernel with a stride of 1, and the number of convolution kernels is 32;

[0084] The fourth residual block contains two convolutional layers, each of which uses a 3×3×3 convolution kernel, and the number of convolution kernels is 128;

[0085] The first, second, and third residual blocks are followed by a maximum pooling layer with a convolution kernel size of 1×1×2;

[0086] The structure of the decoder module is: upsampling-residual block-upsampling-residual block-upsampling-residual block:

[0087] A upsampling layer with a convolution kernel of (1,1,2) is added before each residual block;

[0088] The first residual block contains a concatenation and two convolutional layers. Each convolutional layer uses a 3×3×5 convolution kernel with a stride of 1, and the number of convolution kernels is 32.

[0089] The second residual block contains a connection and two convolutional layers. Each convolutional layer uses a 3×3×5 convolution kernel with a stride of 1, and the number of convolution kernels is 16.

[0090] The third residual block contains a connection and two convolutional layers. Each convolutional layer uses a 3×3×7 convolution kernel with a stride of 1, and the number of convolution kernels is 8;

[0091] The last convolution layer uses a 1×1×1 convolution kernel with a stride of 1 and a number of convolution kernels of 1;

[0092] Each convolution layer uses 3D convolution. To avoid block effects, reflection padding is used before each 3D convolution. And the ReLU activation function is used after each 3D convolution.

[0093] Step 15: Input the training data pair in step 13 into the reconstruction network model in step 14, training data y1, y2, …, y i By reconstructing each convolutional layer of the network model respectively: making the training data y1,y2,…,y i Restore low-dimensional measurement data to high-dimensional reconstructed data

[0094]

[0095] Step 16: Calculate the error between the reconstructed data and the original hyperspectral data through the loss function and perform back propagation to update the parameters of the reconstruction network model in step 14;

[0096] Input training data and train the reconstructed network model in the Keras framework in a supervised manner;

[0097] The loss function uses the mean square error function MSE to calculate the loss value of each input hyperspectral data and its corresponding reconstructed image;

[0098] The total loss value is calculated and optimized using the ADAM optimizer with a learning rate of 0.001.

[0099] Step 17: Repeat steps 15-16 until the error calculated by the loss function reaches a satisfactory value, and then end the training. In this example, the specific hyperparameters used are as follows: total number of training rounds (epoch): 38, number of samples in one training (batch size): 2.

[0100] In the present invention, the reconstruction result is evaluated by using the peak signal-to-noise ratio (PSNR), a commonly used indicator in the field of image quality evaluation, to evaluate the accuracy of the reconstructed data:

[0101]

[0102] Among them, MSE represents the root mean square error between the original image and the reconstructed image. The larger the PSNR index value, the better the reconstruction effect.

[0103] Step 18: Select M original hyperspectral data from the original hyperspectral data set in step 12 as the test set, where M = S, that is, the test set and the validation set contain the same amount of hyperspectral data; use the measurement matrix Φ to compress and sample the original hyperspectral data in the test set to obtain test compressed data sets; use the block processing method in step 13 to block the data in the test set and the test compressed data set to form test data; input the test data into the trained reconstruction network model obtained in step 17 for testing. If the test is qualified, the trained reconstruction network model is put into use, otherwise repeat steps 11-18.

[0104] Specifically, step 2 includes the following steps:

[0105] Step 21: compress and sample the hyperspectral data using the measurement matrix Φ to obtain compressed hyperspectral data Y; the measurement matrix Φ adopts the measurement matrix Φ designed in step 1;

[0106] Step 22: Divide the compressed hyperspectral data Y into blocks to obtain Y1, Y2, …, Y i ; The block processing method is the same as the block processing method of step 13 in step 1;

[0107] Step 23: Use the trained reconstruction network model to reconstruct Y1, Y2, …, Y i Reconstruction is performed, and the reconstructed image blocks are rearranged in sequence to obtain a reconstructed image.

[0108] In the present invention, the first layer in the reconstruction network uses a pseudo-inverse operation to generate preliminary initial reconstruction data, which is more conducive to obtaining a reconstruction effect close to the original data.

[0109] In the present invention, a convolutional neural network is used to solve the problem of compressed sensing hyperspectral data reconstruction. The network is trained to learn the inverse transformation from compressed measurement data to original data, and the trained model is further used to reconstruct the compressed hyperspectral data. This achieves fast and accurate reconstruction of compressed sensing hyperspectral data. Compared with traditional iterative reconstruction algorithms, the reconstruction quality is improved, and the calculation time for reconstructing hyperspectral data is significantly improved, which is much faster than traditional algorithms.

[0110] The present invention is described above by way of example in conjunction with the accompanying drawings. It is obvious that the implementation of the present invention is not limited to the above-mentioned method. As long as various improvements are made using the method concept and technical solution of the present invention, or the concept and technical solution of the present invention are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. A method for reconstructing compressed sensing hyperspectral data based on deep learning, characterized in that: The following steps are involved: Step 1: Train the compressed sensing hyperspectral data reconstruction network model based on convolutional neural network; Step 11: Design the measurement matrix Φ; Step 12: Select an original hyperspectral data set, use K original hyperspectral data as the training set, and select S original hyperspectral data as the validation set; Use the measurement matrix Φ to compress and sample the original hyperspectral data x in the training set and the validation set to obtain the compressed data y; Step 13: Block processing is performed on the compressed data y and the original hyperspectral data x respectively; Raw hyperspectral data Divide into a set of small cubes x1, x2, ..., x i ,in L x ×L y and N λ are the spatial resolution and number of bands of the original hyperspectral data, N x ×N y is the spatial resolution of the small cube; Likewise, the compressed data Divide into y1,y2,...,y i ,in M λ is the number of bands of compressed data; compressed data block y1,y2,...,y i With the original hyperspectral data block x1,x2,...,x i One-to-one correspondence forms training data pairs, making each size N x ×N y ×M λ Each compressed data block has its corresponding true value N x ×N y ×N λ ; Step 14: Construct a compressed sensing hyperspectral data reconstruction network model; The step 14 comprises: step 141: obtaining initial reconstructed data by pseudo-inverse operation; the function of pseudo-inverse operation is to reversely project the hyperspectral data from the compression domain to the spectral domain; The pseudo-inverse operation is performed using N λ This is achieved by a 2D convolutional layer with a convolution kernel size of 1×1. This operation generates the initial reconstructed data, and the weight of the convolutional layer is determined by the pseudo-inverse matrix. The calculation process of the pseudo-inverse matrix is: Setting this layer to non-trainable during network training can avoid overfitting; Step 142: The initial reconstructed data is subjected to a Vnet model to extract features; Vnet includes an encoder and a decoder, and a connection is added between the encoder and the decoder: The structure of the encoder module is: residual block-downsampling-residual block-downsampling-residual block-downsampling-residual block; The first residual block contains 2 convolutional layers, each of which uses a 3×3×7 convolution kernel with a stride of 1, and the number of convolution kernels is 8; The second residual block contains 2 convolutional layers, each of which uses a 3×3×5 convolution kernel with a stride of 1, and the number of convolution kernels is 16; The third residual block contains 2 convolutional layers, each of which uses a 3×3×5 convolution kernel with a stride of 1, and the number of convolution kernels is 32; The fourth residual block contains two convolutional layers, each of which uses a 3×3×3 convolution kernel, and the number of convolution kernels is 128; The first, second, and third residual blocks are followed by a maximum pooling layer with a convolution kernel size of 1×1×2; The structure of the decoder module is: upsampling-residual block-upsampling-residual block-upsampling-residual block: A upsampling layer with a convolution kernel of (1,1,2) is added before each residual block; The first residual block contains a concatenation and two convolutional layers. Each convolutional layer uses a 3×3×5 convolution kernel with a stride of 1, and the number of convolution kernels is 32. The second residual block contains a concatenation and two convolutional layers. Each convolutional layer uses a 3×3×5 convolution kernel with a stride of 1, and the number of convolution kernels is 16. The third residual block contains a concatenation and two convolutional layers. Each convolutional layer uses a 3×3×7 convolution kernel with a stride of 1, and the number of convolution kernels is 8. The last convolution layer uses a 1×1×1 convolution kernel with a stride of 1 and a number of convolution kernels of 1; Each convolution layer uses 3D convolution. To avoid blocking effects, reflection padding is used before each 3D convolution, and the ReLU activation function is used after each 3D convolution. Step 15: Input the training data pairs in step 13 into the reconstruction network model in step 14, and compress the data blocks y1, y2, ..., y i By reconstructing each convolutional layer of the network model, the compressed data blocks y1, y2, ..., y i Recovering from low-dimensional measurement data to high-dimensional reconstructed data Step 16: Calculate the error between the reconstructed data and the original hyperspectral data through the loss function and perform back propagation to update the parameters of the reconstruction network model in step 14; Step 17: Repeat steps 15-16 until the error calculated by the loss function reaches a satisfactory value, and then end the training. At this time, the trained reconstruction network model is obtained; Step 2: Use the reconstruction network model trained in step 1 to reconstruct the compressed sensing hyperspectral data.

2. The method for reconstructing compressed sensing hyperspectral data based on deep learning according to claim 1, characterized in that: The step 1 also includes step 18: In the original hyperspectral data set in step 12, M original hyperspectral data are selected as the test set, where M=S, that is, the test set and the validation set contain the same number of hyperspectral data; the original hyperspectral data in the test set are compressed and sampled using the measurement matrix Φ to obtain the test compressed data set; The block processing method in step 13 is used to block the data in the test set and the test compressed data set to form test data; The test data are respectively input into the trained reconstructed network model obtained in step 17 for testing. If the test is qualified, the trained reconstructed network model is put into use. Otherwise, steps 11-18 are repeated.

3. The method for reconstructing compressed sensing hyperspectral data based on deep learning according to claim 1 or 2, characterized in that: In step 11, the measurement matrix Φ adopts a random Gaussian measurement matrix or a Bernoulli matrix.

4. The method for reconstructing compressed sensing hyperspectral data based on deep learning according to claim 1 or 2, characterized in that: In step 12, the hyperspectral data in the original hyperspectral data set is the original hyperspectral data after spatial downsampling processing.

5. The method for reconstructing compressed sensing hyperspectral data based on deep learning according to claim 1 or 2, characterized in that: In step 12, only the spectral dimension of the original hyperspectral data x is compressed, and the spatial resolution remains unchanged. The processing method of the compressed sampling of the original hyperspectral data x is y=Φx, where the measurement matrix is 6. The method for reconstructing compressed sensing hyperspectral data based on deep learning according to claim 1 or 2, characterized in that: In step 16, the loss function uses the mean square error function MSE to calculate the loss value of each original hyperspectral data and its corresponding reconstructed data; The total loss value is calculated and optimized using the ADAM optimizer with a learning rate of 0.

001.

7. The method for reconstructing compressed sensing hyperspectral data based on deep learning according to claim 1 or 2, characterized in that: In step 17, the reconstruction result is evaluated using the peak signal-to-noise ratio (PSNR) to assess the accuracy of the reconstructed data: Among them, MSE represents the mean square error between the original image and the reconstructed image. The larger the PSNR index value, the better the reconstruction effect. n represents the quantization bit depth of the image, and n is generally a positive integer.

8. The method for reconstructing compressed sensing hyperspectral data based on deep learning according to claim 1, characterized in that: The step 2 comprises: Step 21: compress and sample the hyperspectral data using the measurement matrix Φ to obtain compressed hyperspectral data Y; the measurement matrix Φ adopts the measurement matrix Φ designed in step 1; Step 22: Divide the compressed hyperspectral data Y into blocks to obtain Y1, Y2, ..., Y i ; The block processing method is the same as the block processing method of step 13 in step 1; Step 23: Use the trained reconstruction network model to reconstruct Y1,Y2,...,Y i Reconstruction is performed, and the reconstructed image blocks are rearranged in sequence to obtain a reconstructed image.

Images