Hyperspectral information acquisition transmission method and equipment against atmospheric turbulence

Abstract

This application provides a method and device for acquiring and transmitting hyperspectral information with resistance to atmospheric turbulence. The method includes: encoding and compressing a hyperspectral image using a diffractive optical element to obtain an RGB image; inputting the RGB image into an encoding neural network, which encodes and compresses the image and outputs a one-dimensional feature stream; transmitting the stream to a receiver via an FSO channel, where the receiver arranges the stream into a feature map of a preset shape, inputs the feature map into a decoding neural network, and decodes the image to output a reconstructed hyperspectral image; wherein the diffractive optical element, the encoding neural network, and the decoding neural network are obtained through end-to-end joint training under a channel model that includes atmospheric turbulence noise. This application enables the integrated acquisition and transmission of large-scale hyperspectral remote sensing information from satellite and ground, saving bandwidth resources, improving robustness to channel impairments, increasing transmission capacity, and improving interpretation quality.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a method and device for acquiring and transmitting hyperspectral information that is resistant to atmospheric turbulence. Background Technology

[0002] As a crucial component of space remote sensing technology, hyperspectral remote sensing, with its unique spectral detection capabilities, has been widely applied in fields such as national defense, environmental monitoring, mineral exploration, and agricultural monitoring. However, with the rapid growth of space exploration operations, the volume and data flow of hyperspectral remote sensing services have surged. Current systems are struggling to support the high-speed acquisition and transmission of massive amounts of hyperspectral data, severely impacting real-time analysis and decision-making in remote sensing operations. Simultaneously, atmospheric turbulence between space and ground significantly affects communication quality, further reducing the transmission capacity and rate of hyperspectral remote sensing information. To achieve ultra-high-speed, high-capacity, and channel-resistant real-time acquisition and transmission of hyperspectral remote sensing information, an integrated system combining diffractive optical element (DOE) coded hyperspectral acquisition technology and free-space optical communication technology has been proposed. Currently, advanced hyperspectral imaging systems based on DOE encoding can achieve ultra-high-speed hyperspectral imaging.

[0003] However, in hyperspectral remote sensing scenarios, existing diffraction-coded hyperspectral imaging systems and algorithms cannot cope with the damage caused by free-space optical (FSO) channel noise, resulting in degraded quality of the hyperspectral information recovered after satellite-to-ground transmission. Furthermore, existing satellite-to-ground communication technologies are only suitable for encoding and transmitting natural images; when encoding optical measurement maps acquired by DOE spectral imaging systems, additional coding noise is introduced, further reducing the quality of the final recovered hyperspectral information. Summary of the Invention

[0004] In view of this, embodiments of this application provide a hyperspectral information acquisition and transmission method and device that resists atmospheric turbulence, so as to eliminate or improve one or more defects existing in the prior art.

[0005] One aspect of this application provides a method for acquiring and transmitting hyperspectral information resistant to atmospheric turbulence, the method comprising the following steps: The hyperspectral image is encoded and compressed using a preset diffractive optical element to obtain a three-channel RGB image corresponding to the hyperspectral image; The three-channel RGB image is input into a preset encoding neural network, which performs normalization, convolution, downsampling, dense connection, importance-aware weighting, and compression on the three-channel RGB image, and outputs a one-dimensional feature stream. Based on a preset FSO channel, the one-dimensional feature stream is transmitted to the receiving end, so that the receiving end rearranges the received one-dimensional feature stream into a feature map of a preset shape, and inputs the feature map of the preset shape into a preset decoding neural network, and the decoding neural network performs decompression, upsampling, dense connection, convolution and denormalization processing on the feature map of the preset shape, and outputs the reconstructed hyperspectral image corresponding to the hyperspectral image; wherein, the diffractive optical element, the encoding neural network and the decoding neural network are pre-trained in a channel model containing atmospheric turbulence noise based on an end-to-end joint training method.

[0006] In some embodiments of this application, the encoded neural network includes: The normalization layer is used to normalize the three-channel RGB image to obtain the corresponding normalized three-channel RGB image; The first convolutional layer is used to increase the number of channels of the normalized three-channel RGB image to a preset first number of channels to obtain the first convolutional feature map; Multiple cascaded downsampling layers are used to downsample the first convolutional feature map to obtain a downsampled feature map with a preset number of second channels; wherein, the downsampling operation includes halving the spatial size of the feature map and doubling the number of channels; The first dense connection layer is used to copy the feature maps before entering each of the downsampling layers and convolve them based on a plurality of branch convolutional layers of different preset lengths to obtain branch feature maps corresponding to each of the branch convolutional layers. The branch feature maps are then added to the downsampling feature maps to obtain the first dense connection feature map. The first dense connection layer includes the same number of branch convolutional layers as the downsampling layers. The importance-aware layer is used to generate corresponding weights for the first densely connected feature maps of each channel, and perform channel-by-channel weighting to obtain the importance-aware feature maps. The compression layer is used to compress the size of the importance-aware feature map to the target dimension according to a preset compression ratio, and flatten the compressed importance-aware feature map to output the one-dimensional feature code stream.

[0007] In some embodiments of this application, the decoding neural network includes: The decompression layer is used to decompress the received feature map of the preset shape to obtain the decompressed feature map; Multiple cascaded upsampling layers are used to upsample the decompressed feature map to obtain an upsampled feature map; wherein, the upsampling operation includes doubling the spatial size of the feature map and halving the number of channels to restore the spatial resolution of the feature map; The second dense connection layer is used to copy the feature maps before entering each of the upsampling layers and convolve them based on multiple pre-set transposed convolutional layers of different lengths to obtain transposed convolutional feature maps corresponding to each of the transposed convolutional layers. The transposed convolutional feature maps are then added to the upsampling feature maps to obtain the second dense connection feature map. The second dense connection layer includes the same number of transposed convolutional layers as the upsampling layers. The second convolutional layer is used to adjust the number of channels of the second densely connected feature map to a preset third number of channels to obtain the second convolutional feature map; The denormalization layer is used to denormalize the second convolutional feature map and output the reconstructed hyperspectral image.

[0008] In some embodiments of this application, the end-to-end joint training specifically includes: The first update step is performed as follows: a physical height map of the diffractive optical element is generated based on the current trainable parameter map; the point spread function at different wavelengths is calculated based on the physical height map; the original hyperspectral image is convolved with and superimposed with the point spread functions of each wavelength to generate a two-dimensional coded image; the two-dimensional coded image is sequentially encoded by a coding neural network, transmitted through a channel model containing atmospheric turbulence noise, and decoded by a decoding neural network to obtain the recovered hyperspectral image; the mean square error between the recovered hyperspectral image and the original hyperspectral image is calculated as the loss function. The second update step is performed: the gradient of the loss function is propagated back to the trainable parameter map of the diffractive optical element by passing the decoding neural network, the channel model, and the encoding neural network in sequence through the gradient descent algorithm, so as to synchronously update the parameters of the diffractive optical element, the weights of the encoding neural network, and the weights of the decoding neural network. Determine whether the loss function has converged; if it has not converged, continue to execute the first update step and the second update step; until the loss function converges, use the updated diffractive optical element, encoding neural network and decoding neural network as the diffractive optical element, encoding neural network and decoding neural network for hyperspectral information processing.

[0009] In some embodiments of this application, the end-to-end joint training further includes: Perform the neural network update step: fix the parameters of the synchronously updated diffractive optical element, and update the weights of the decoding neural network based on the added atmospheric turbulence noise and the loss function. Determine whether the loss function has converged; if it has not converged, continue to execute the first update step and the neural network update step; until the loss function converges, and use the updated encoding neural network and decoding neural network as the trained encoding neural network and decoding neural network; Perform the diffractive optical element update step: fix the weights of the decoding neural network obtained from the neural network update step, and update the parameters of the diffractive optical element based on the added atmospheric turbulence noise and the loss function; Determine whether the loss function has converged; if it has not converged, continue to execute the first update step and the diffractive optical element update step; until the loss function converges, and use the updated diffractive optical element as the trained diffractive optical element.

[0010] In some embodiments of this application, the hyperspectral image includes information from different spectral bands; Correspondingly, the step of encoding and compressing the hyperspectral image using a preset diffraction optical element to obtain a three-channel RGB image corresponding to the hyperspectral data cube includes: The information of different bands in the hyperspectral image is encoded using a preset diffraction optical element to generate a point spread function corresponding to the information of each band. Based on each of the point spread functions, the hyperspectral image is compressed using a preset RGB camera to obtain the three-channel RGB image.

[0011] In some embodiments of this application, transmitting the one-dimensional feature stream to the receiving end via a preset FSO channel includes: The one-dimensional feature code stream is modulated onto an optical carrier to generate an optical signal carrying the one-dimensional feature code stream information; and the optical signal is transmitted to the receiving end through the FSO optical channel so that the receiving end can perform photoelectric conversion and sampling on the received optical signal to obtain the recovered one-dimensional feature code stream.

[0012] Another aspect of this application provides an electronic device, including a processor and a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned method for acquiring and transmitting hyperspectral information against atmospheric turbulence.

[0013] A third aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for acquiring and transmitting hyperspectral information against atmospheric turbulence.

[0014] The fourth aspect of this application provides a computer program product comprising a computer program that, when executed by a processor, implements the aforementioned method for acquiring and transmitting hyperspectral information against atmospheric turbulence.

[0015] This application discloses a hyperspectral information acquisition and transmission method resistant to atmospheric turbulence. The method includes the following steps: encoding and compressing a hyperspectral image using a pre-defined diffractive optical element to obtain a three-channel RGB image corresponding to the hyperspectral image; inputting the three-channel RGB image into a pre-defined encoding neural network, which performs normalization, convolution, downsampling, dense connection, importance-aware weighting, and compression processing on the three-channel RGB image, and outputs a one-dimensional feature stream; transmitting the one-dimensional feature stream to a receiving end based on a pre-defined FSO channel, so that the receiving end rearranges the received one-dimensional feature stream into a feature map of a pre-defined shape, inputs the feature map of the pre-defined shape into a pre-defined decoding neural network, and performs decompression, upsampling, dense connection, convolution, and de-normalization processing on the one-dimensional feature stream, outputting a reconstructed hyperspectral image corresponding to the hyperspectral image; wherein the diffractive optical element, the encoding neural network, and the decoding neural network are pre-trained in an end-to-end joint training manner under a channel model containing atmospheric turbulence noise. It can realize the integrated acquisition and transmission of large-scale hyperspectral remote sensing information from space and ground, while saving bandwidth resources, resisting FSO channel noise damage, improving robustness to channel damage, increasing transmission capacity and improving interpretation quality.

[0016] Additional advantages, objectives, and features of this application will be set forth in part in the description which follows, and will in part become apparent to those skilled in the art upon review of the following description, or may be learned by practice of the application. The objectives and other advantages of this application can be realized and obtained by means of the structures specifically pointed out in the specification and drawings.

[0017] Those skilled in the art will understand that the purposes and advantages that can be achieved with this application are not limited to those specifically described above, and that the above and other purposes that this application can achieve will be more clearly understood from the following detailed description. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, do not constitute a limitation thereof. The components in the drawings are not drawn to scale but are merely for illustrating the principles of this application. For ease of illustration and description of certain parts of this application, corresponding portions in the drawings may be enlarged, i.e., may appear larger relative to other components in an exemplary device actually manufactured according to this application. In the drawings: Figure 1 This is a schematic diagram of the first process of a hyperspectral information acquisition and transmission method for resisting atmospheric turbulence in one embodiment of this application.

[0019] Figure 2This is a schematic diagram of the second process of a hyperspectral information acquisition and transmission method for resisting atmospheric turbulence in one embodiment of this application.

[0020] Figure 3 This is a schematic diagram of the third process of a hyperspectral information acquisition and transmission method against atmospheric turbulence in one embodiment of this application.

[0021] Figure 4 This is a schematic diagram of an integrated system structure for hyperspectral information acquisition and transmission that resists atmospheric turbulence, as described in a specific example of the hyperspectral information acquisition and transmission method for resisting atmospheric turbulence in this application.

[0022] Figure 5 This is a schematic diagram of a neural network encoding and decoding network structure for a hyperspectral information acquisition and transmission method that resists atmospheric turbulence, as described in a specific example of this application.

[0023] Figure 6 This is a schematic diagram of a DOE generation process for a hyperspectral information acquisition and transmission method that resists atmospheric turbulence, as illustrated in a specific example of this application.

[0024] Figure 7 This is a schematic diagram of an end-to-end joint training process for a hyperspectral information acquisition and transmission method that resists atmospheric turbulence, as illustrated in a specific example of this application.

[0025] Figure 8 This is a performance comparison chart of a hyperspectral information acquisition and transmission method for resisting atmospheric turbulence in a specific example of this application, using and not using a three-stage training strategy.

[0026] Figure 9 This is a schematic diagram of a direct-modulation and direct-detection FSO communication system, which is a specific example of a hyperspectral information acquisition and transmission method resistant to atmospheric turbulence in this application.

[0027] Figure 10 This is a comparison chart of the performance indicators of hyperspectral acquisition and transmission under different turbulence intensities in a specific example of the hyperspectral information acquisition and transmission method for resisting atmospheric turbulence in this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and their descriptions are used to explain this application, but are not intended to limit it.

[0029] It should also be noted that, in order to avoid obscuring this application with unnecessary details, only the structures and / or processing steps closely related to the solution according to this application are shown in the accompanying drawings, while other details that are not closely related to this application are omitted.

[0030] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0031] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0032] In the following description, embodiments of the present application will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0033] It should be noted that existing DOE-based hyperspectral imaging systems can achieve ultra-high-speed hyperspectral imaging. These systems use diffractive optical coding elements to perform unique phase coding on spectral information in different bands, and finally use an RGB camera to capture a single RGB compressed measurement image. Algorithms are then used to achieve high-precision and rapid hyperspectral information recovery. In existing satellite-to-ground communication technologies, source coding (such as JPEG2000) is designed for natural images, and its statistical assumptions and redundancy removal methods match those of natural images. However, DOE output images are optically encoded modulated images with pixel grayscale distributions and spatial correlations completely different from natural images. Directly applying traditional coding methods introduces additional noise and distortion, destroying the encoded information in the encoded image and leading to a decrease in reconstruction quality. In other words, when using DOE for hyperspectral remote sensing, the system and algorithm cannot cope with the damage caused by FSO channel noise, resulting in poor quality of hyperspectral information recovered from satellite-to-ground transmission. Existing satellite-to-ground communication technologies are only suitable for encoding and transmitting natural images; encoding the optical measurement images acquired by the DOE spectral imaging system introduces additional coding noise, further reducing the final quality of hyperspectral information recovery. Based on the above problems and the challenges posed by the surge in hyperspectral remote sensing traffic and data flow, the inventors of this application first conceived of using a DOE-coded hyperspectral imaging system for hyperspectral information acquisition. Utilizing convolutional neural networks from machine learning, joint coding and modulation, and joint decoding and interpretation are implemented at both the transmitting and receiving ends. Furthermore, this application can perform end-to-end training with the FSO channel and neural network, directly learning the data distribution characteristics of the coded image. It learns specific feature extraction and compressed representation based on the statistical laws of the DOE-coded image, enabling the network to optimize source-channel coding while maintaining the integrity of spectral information. This fundamentally eliminates the adaptation noise introduced by traditional coding. Furthermore, by updating the structural parameters of the DOE and the parameters of the neural network, robustness to channel impairments can be improved. This enables integrated acquisition and transmission of large-scale hyperspectral remote sensing information from both satellite and ground, while saving bandwidth resources, resisting FSO channel noise impairments, and offering advantages such as large transmission capacity and high interpretation quality.

[0034] The following examples will provide a detailed description.

[0035] This application provides a method for acquiring and transmitting hyperspectral information that resists atmospheric turbulence. See also... Figure 1 The method includes the following steps: Step 100: Encode and compress the hyperspectral image using a preset diffraction optical element to obtain a three-channel RGB image corresponding to the hyperspectral image; In step 100, the hyperspectral image can be a hyperspectral data cube with a wavelength range of 400-700 nm, a spatial resolution of 512×512, and 31 bands, with pixel values ​​ranging from 0 to 65536; the size of the three-channel RGB image can be 3×512×512. For 31-band hyperspectral data, traditional hyperspectral imaging requires acquiring spectral images of all bands, with the data volume increasing linearly with the number of bands. However, the DOE system can compress the 31-band hyperspectral image into RGB data of 3 bands, directly reducing the amount of data acquired by 10 times, which can directly alleviate the pressure on on-board storage writes and the downlink transmission bandwidth requirements.

[0036] Step 200: Input the three-channel RGB image into a preset encoding neural network, so that the encoding neural network performs normalization, convolution, downsampling, dense connection, importance-aware weighting and compression processing on the three-channel RGB image, and outputs a one-dimensional feature code stream; In step 200, the three-channel RGB image is input into the first normalization layer of the encoding neural network, linearly mapping the pixel values ​​from the original range of 0~65535 to the interval [0,1] to obtain a normalized three-channel RGB image with a data size of 3×512×512. The first convolutional layer of the encoding neural network convolves the normalized three-channel RGB image to obtain a first convolutional feature map with a data size of 32×512×512. The first convolutional feature map is passed sequentially through three cascaded downsampling layers. Each downsampling operation halves the spatial size of the feature map (height and width are each divided by 2) and doubles the number of channels. After each downsampling layer, a dense connection operation is performed, copying the current feature map and inputting it into three independent convolutional layers. These three convolutional layers have a kernel size of 3, 256 kernels, and strides of 8, 4, and 2 respectively, generating three branch feature maps. The data size of each branch feature map is adjusted to 256×64×64. Then, the three branch feature maps are added element-wise to the current feature map of the main branch to obtain the enhanced feature map. After passing through the three cascaded downsampling layers, the first densely connected feature map is obtained, with a data size of 256×64×64. The first densely connected feature map is then passed through an importance-aware layer. This layer generates a learnable weight coefficient for each channel (256 channels in total). The feature map of each channel is multiplied by the corresponding weight coefficient to obtain the importance-aware feature map, which also has a data size of 256×64×64. According to a preset compression ratio γ, the importance-aware feature map is processed by a compression layer, which transforms the size of the importance-aware feature map into a compressed importance-aware feature map of γ×32×32, where γ can be a positive integer less than 256. The compressed importance-aware feature map is then flattened into a one-dimensional vector to obtain a one-dimensional feature stream of length 4096×γ, which is then output to the FSO channel.

[0037] Step 300: Based on a preset FSO channel, the one-dimensional feature stream is transmitted to the receiving end, so that the receiving end rearranges the received one-dimensional feature stream into a feature map of a preset shape, and inputs the feature map of the preset shape into a preset decoding neural network, and the decoding neural network performs decompression, upsampling, dense connection, convolution and denormalization processing on the feature map of the preset shape, and outputs the reconstructed hyperspectral image corresponding to the hyperspectral image; wherein, the diffractive optical element, the encoding neural network and the decoding neural network are pre-trained in a channel model containing atmospheric turbulence noise based on an end-to-end joint training method.

[0038] In step 300, the receiving end rearranges the received one-dimensional feature stream into a γ×32×32 received feature map according to the inverse transformation rule corresponding to the compression layer of the sending end. The received feature map is then input into the decoding neural network, so that the decompression layer in the receiving end transforms the feature map size into a 256×64×64 decompressed feature map through a transpose convolution operation. The decompressed feature map is sequentially passed through three cascaded upsampling layers. Each upsampling operation doubles the spatial size of the feature map (height and width are each multiplied by 2) and halves the number of channels. The residual convolutional module in each upsampling layer does not change the feature map size. After each upsampling layer, a dense connection operation is performed, copying the current feature map and inputting it into three independent transposed convolutional layers. These three transposed convolutional layers all have a kernel size of 3, a kernel count of 32, and strides of 8, 4, and 2 respectively, generating three transposed convolutional feature maps. The data size of all three transposed convolutional feature maps is adjusted to 32×512×512. Then, the three transposed convolutional feature maps are added element-wise to the upsampling feature map to obtain a second densely connected feature map with a data size of 32×512×512. The second densely connected feature map is then passed through a second convolutional layer to obtain a second convolutional feature map with a data size of 31×512×512. The second convolutional feature map is denormalized so that the pixel values ​​are linearly mapped back from the [0,1] interval to the original range of 0~65535, thus obtaining a reconstructed hyperspectral image. The data size of the reconstructed hyperspectral image is 31×512×512, and the value range is 0~65535.

[0039] As described above, the hyperspectral information acquisition and transmission method against atmospheric turbulence provided in this application acquires hyperspectral information through a DOE-encoded hyperspectral imaging system, achieving high-speed acquisition, saving bandwidth resources, and increasing transmission capacity. It utilizes convolutional neural networks from machine learning to achieve joint coding modulation and joint decoding at both the transmitting and receiving ends. Furthermore, it performs end-to-end optimization using the FSO channel and improves robustness to channel impairments by updating the DOE's structural parameters and the neural network's parameters, thereby enhancing hyperspectral information extraction and compression capabilities. Ultimately, it enables integrated acquisition and transmission of large-scale hyperspectral remote sensing information from satellite and ground, resisting FSO channel noise impairments, ensuring uninterrupted link operation, providing relatively stable communication quality, and improving the encoding and decoding efficiency of the coding neural network, thus enhancing interpretation quality.

[0040] To further improve the coding efficiency of the coding neural network, in the hyperspectral information acquisition and transmission method against atmospheric turbulence provided in this application embodiment, the coding neural network includes: The normalization layer is used to normalize the three-channel RGB image to obtain the corresponding normalized three-channel RGB image; The first convolutional layer is used to increase the number of channels of the normalized three-channel RGB image to a preset first number of channels to obtain the first convolutional feature map; Multiple cascaded downsampling layers are used to downsample the first convolutional feature map to obtain a downsampled feature map with a preset number of second channels; wherein, the downsampling operation includes halving the spatial size of the feature map and doubling the number of channels; The first dense connection layer is used to copy the feature maps before entering each of the downsampling layers and convolve them based on a plurality of branch convolutional layers of different preset lengths to obtain branch feature maps corresponding to each of the branch convolutional layers. The branch feature maps are then added to the downsampling feature maps to obtain the first dense connection feature map. The first dense connection layer includes the same number of branch convolutional layers as the downsampling layers. The importance-aware layer is used to generate corresponding weights for the first densely connected feature maps of each channel, and perform channel-by-channel weighting to obtain the importance-aware feature maps. The compression layer is used to compress the size of the importance-aware feature map to the target dimension according to a preset compression ratio, and flatten the compressed importance-aware feature map to output the one-dimensional feature code stream.

[0041] In one or more embodiments of this application, the numerical range of the three-channel RGB image after normalization by the normalization layer is 0 to 1. After convolution by the first convolutional layer, a first convolutional feature map of size 32×512×512 is obtained. Multiple cascaded downsampling layers in the encoding process of the coding neural network can double the first dimension of the first convolutional feature map and halve the last two dimensions. These multiple cascaded downsampling layers can be three cascaded downsampling layers; the residual convolution in these multiple cascaded downsampling layers is used to perform feature transformation on the feature map without changing the data shape. After three cascaded downsampling layers, a downsampled feature map of size 256×64×64 is obtained. In the encoding process of the coding neural network, the first dense connection layer can improve the efficiency of network encoding. The specific process of dense connection processing is that before each downsampling layer, the data is copied and enters an additional convolutional layer. After passing through the convolutional layer, a branch feature map of size 256×64×64 is generated. From beginning to end, each convolutional operation in the dense connection uses a kernel size of 3, a kernel count of 256, and strides of 8, 4, and 2 respectively. Finally, the branch feature maps generated from the three branches are added to the downsampled feature map and then processed by the importance-aware layer. The importance-aware layer generates 256 weights and multiplies each weight by 256 64×64 branch feature maps, redistributing feature importance without changing the data size, resulting in an importance-aware feature map. Next, the compression layer compresses the data size to a target dimension according to a pre-set compression ratio, where the target dimension can be γ×32×32, and γ is the compression ratio. The compressed importance-aware feature map is then flattened into a one-dimensional feature stream of length 4096γ and transmitted through the FSO channel.

[0042] To further improve the decoding efficiency of the decoding neural network, in the hyperspectral information acquisition and transmission method against atmospheric turbulence provided in this application embodiment, the decoding neural network includes: The decompression layer is used to decompress the received feature map of the preset shape to obtain the decompressed feature map; Multiple cascaded upsampling layers are used to upsample the decompressed feature map to obtain an upsampled feature map; wherein, the upsampling operation includes doubling the spatial size of the feature map and halving the number of channels to restore the spatial resolution of the feature map; The second dense connection layer is used to copy the feature maps before entering each of the upsampling layers and convolve them based on multiple pre-set transposed convolutional layers of different lengths to obtain transposed convolutional feature maps corresponding to each of the transposed convolutional layers. The transposed convolutional feature maps are then added to the upsampling feature maps to obtain the second dense connection feature map. The second dense connection layer includes the same number of transposed convolutional layers as the upsampling layers. The second convolutional layer is used to adjust the number of channels of the second densely connected feature map to a preset third number of channels to obtain the second convolutional feature map; The denormalization layer is used to denormalize the second convolutional feature map and output the reconstructed hyperspectral image.

[0043] In one or more embodiments of this application, the feature map with the preset shape can be a feature map with a data size of γ×32×32. The rearranged feature map first passes through a decompression layer in the interpreting neural network to generate a decompressed feature map of size 256×64×64. Multiple cascaded upsampling layers can be three cascaded upsampling layers; the three cascaded upsampling layers halve the first dimension of the decompressed feature map and double the second and third dimensions, generating an upsampled feature map with a data size of 32×512×512. Finally, after passing through a second convolutional layer and a denormalization layer, a reconstructed hyperspectral image of size 31×512×512 is restored, and the numerical range of the reconstructed hyperspectral image is from 0 to 65536. In the neural network interpretation process, a dense connection processing method is also used. Before each pass through the upsampling layer, the data is copied and passed through a transposed convolutional layer. Each branch generates data of size 32×512×512. Each transposed convolutional layer has a kernel size of 3 and a kernel count of 32, with strides of 8, 4, and 2 respectively from front to back. Finally, all branch data are added to the main branch data before entering the second convolutional layer.

[0044] To further enhance the system's robustness to FSO channel noise and thus achieve optimal hyperspectral information extraction and compression capabilities, this application provides a hyperspectral information acquisition and transmission method resistant to atmospheric turbulence, see [reference needed]. Figure 2 The end-to-end joint training specifically includes: Step 010: Perform the first update step: Generate a physical height map of the diffractive optical element based on the current trainable parameter map; calculate the point spread function at different wavelengths based on the physical height map; convolve and superimpose the original hyperspectral image with the point spread function of each wavelength to generate a two-dimensional coded image; encode the two-dimensional coded image sequentially through a coding neural network, transmit it through a channel model containing atmospheric turbulence noise, and decode it through a decoding neural network to obtain the recovered hyperspectral image; calculate the mean square error between the recovered hyperspectral image and the original hyperspectral image as the loss function. In step 010, a 1024×1024 trainable parameter map is initialized. Based on this parameter map, a physical height map of the DOE is generated, and the point spread function at different wavelengths is calculated. The hyperspectral data cube is convolved with and superimposed with the point spread functions for each wavelength, simulating the DOE's image encoding process to generate a two-dimensional encoded image. The generated two-dimensional encoded image is then processed through neural network encoding, channel transmission, and neural network decoding to reconstruct the hyperspectral image. The reconstructed hyperspectral image and the original hyperspectral image are used to calculate the mean square error (MSE). The entire process is completely differentiable; therefore, the gradient of the loss function of the reconstructed network can be backpropagated to the 1024×1024 parameter map, and the DOE parameters are directly optimized and updated through gradient descent. The effective modulation region (DOE) structure is generated through the following operations: First, a 1024×1024 floating-point matrix is ​​generated, with each floating-point number ranging from 0 to 1. Then, regions greater than 512 units from the center are set to zero, and all parameters less than or equal to 512 units from the center are set as trainable parameters, thus generating a circular DOE structure with trainable parameters. The physical height map is generated by amplifying the floating-point sequence by a factor of 16 and quantizing it using rounding. All values ​​are quantized to integers, as shown in the following formula: in, These are the quantized parameters; It is the original floating-point sequence; This is a rounding operation. The quantized value is then... Material height converted to DOE The transformation process can be represented by the following formula: in, , , , A physical height map is generated based on the material height of the DOE, and the actual DOE production is then carried out.

[0045] Step 020: Perform the second update step: Using the gradient descent algorithm, the gradient of the loss function is propagated back through the decoding neural network, the channel model, and the encoding neural network to the trainable parameter map of the diffractive optical element, so as to synchronously update the parameters of the diffractive optical element, the weights of the encoding neural network, and the weights of the decoding neural network. In step 020, the update process of the DOE parameters can be represented as follows: in, For the updated DOE parameters; These are the DOE parameters before the update. The learning rate; Indicates gradient calculation; The loss function is MSE; Let E be the decoding neural network; E be the encoding neural network; n be the FSO system noise; X be the original hyperspectral image; Y be the encoded RGB image; and t be the t-th update round. During backpropagation, the gradient of the loss function is propagated sequentially through the decoding network, the channel model, and the encoding network to update the weights of each layer of the neural network. The update process of the neural network parameters can be represented as: in, For the updated network parameters; These are the network parameters before the update.

[0046] Step 030: Determine whether the loss function has converged; if it has not converged, continue to execute the first update step and the second update step; until the loss function converges, use the updated diffractive optical element, encoding neural network and decoding neural network as the diffractive optical element, encoding neural network and decoding neural network for hyperspectral information processing.

[0047] To further enhance the system's robustness to FSO channel noise and thus achieve optimal hyperspectral information extraction and compression capabilities, this application provides a hyperspectral information acquisition and transmission method resistant to atmospheric turbulence, see [reference needed]. Figure 3 The end-to-end joint training also includes: Step 040: Perform the neural network update step: Fix the parameters of the synchronously updated diffractive optical element, and update the weights of the decoding neural network based on the added atmospheric turbulence noise and the loss function. Step 050: Determine whether the loss function has converged; if it has not converged, continue to execute the first update step and the neural network update step; until the loss function converges, and use the updated encoding neural network and decoding neural network as the trained encoding neural network and decoding neural network; Step 060: Perform the diffractive optical element update step: fix the weights of the decoding neural network obtained in the neural network update step, and update the parameters of the diffractive optical element based on the added atmospheric turbulence noise and the loss function. Step 070: Determine whether the loss function has converged; if it has not converged, continue to execute the first update step and the diffractive optical element update step; until the loss function converges, and use the updated diffractive optical element as the trained diffractive optical element.

[0048] In one or more embodiments of this application, the end-to-end joint training is divided into three stages. The first stage is the parameter pre-training stage, in which DOE parameters and network parameters are trained and updated without adding FSO noise, enabling the system to obtain optimal hyperspectral information extraction and compression capabilities. The second stage is the network-channel interaction stage, in which the DOE parameters trained in the first stage are fixed. During the training process, FSO channel noise is added, and the network parameters are updated, allowing the system to initially achieve robustness to FSO channel noise. The third stage is the DOE-channel interaction stage, in which the encoding and decoding network parameters obtained in the second stage are fixed, and the DOE parameters are updated only when FSO channel noise is added.

[0049] To further achieve high-speed acquisition, in a hyperspectral information acquisition and transmission method resistant to atmospheric turbulence provided in this application embodiment, the hyperspectral image includes information from different spectral bands; Correspondingly, step 100 includes: Step 110: Encode the information of different bands in the hyperspectral image using a preset diffraction optical element to generate the point spread function corresponding to the information of each band; Step 120: Based on each of the point spread functions, the hyperspectral image is compressed using a preset RGB camera to obtain the three-channel RGB image.

[0050] In one or more embodiments of this application, diffractive optical elements encode information from different bands in the hyperspectral image and detect it using an RGB camera. A single shot is sufficient to capture all the information of a complete hyperspectral image. In contrast, traditional hyperspectral imaging relies on mechanical scanning or time-division multiplexing acquisition modes that switch band by band, requiring several minutes to acquire a complete hyperspectral image. The DOE system in this application only requires the detector to continuously acquire RGB-encoded images at its own frame rate; the acquisition speed is limited only by the detector's frame rate, not the mechanical scanning speed. Therefore, high-speed acquisition can be achieved, resulting in a significantly higher acquisition speed compared to traditional systems.

[0051] To further overcome atmospheric turbulence signal impairment, ensure uninterrupted link operation, and provide relatively stable communication quality, in a hyperspectral information acquisition and transmission method resistant to atmospheric turbulence provided in this application embodiment, the transmission of the one-dimensional feature code stream to the receiving end based on a preset FSO channel includes: The one-dimensional feature code stream is modulated onto an optical carrier to generate an optical signal carrying the one-dimensional feature code stream information; and the optical signal is transmitted to the receiving end through the FSO optical channel so that the receiving end can perform photoelectric conversion and sampling on the received optical signal to obtain the recovered one-dimensional feature code stream.

[0052] In one or more embodiments of this application, based on a preset arbitrary waveform generator (AWG), the one-dimensional feature stream is converted into an analog driving electrical signal to drive a preset Mach-Zehnder modulator (MZM) to modulate the intensity of a preset wavelength laser generated by a preset laser, thereby obtaining an optical signal. The optical signal is amplified using preset erbium-doped fiber amplifiers (EDFAs), and transmitted to the receiving end via the FSO channel. The receiving end uses a preset photodetector (PD) to convert the received optical signal into a corresponding electrical signal, and samples the electrical signal using a preset digital oscilloscope to obtain the received one-dimensional feature stream. In traditional FSO transmission, turbulence intensity exceeding a threshold causes a surge in the bit error rate, leading to the failure of the source decoding algorithm and the inability to recover image information, i.e., communication interruption. However, the transmission architecture of this application does not rely on bit transmission. When turbulence intensifies, some feature information may be damaged, but the receiver can still infer the approximate content based on the remaining features and the interpretation network. The system will not be interrupted, only the reconstruction accuracy will decrease. The entire end-to-end network is trained in channels containing turbulence of varying intensities, enabling the system to learn how to automatically adjust feature extraction and reconstruction strategies when the channel deteriorates, thus maintaining link connectivity at all times. Therefore, compared to traditional FSO transmission schemes, this application can guarantee uninterrupted link operation and provide relatively stable communication quality.

[0053] In a specific example of the hyperspectral information acquisition and transmission method against atmospheric turbulence in this application, see [reference needed]. Figure 4 and Figure 5 The method includes the following steps: Hyperspectral images with a resolution of 512×512×31, a numerical range of 0 to 65536, and a range of 400 nm to 700 nm are encoded using diffractive optical elements. The hyperspectral images are encoded and compressed into 3-channel RGB images, and the size of the detected RGB images is 3×512×512.

[0054] Next, the RGB image is fed into a neural network for encoding. The specific structures of the encoding neural network and the interpreting neural network are as follows: Figure 5As shown. After the first normalization layer (NL), the numerical range of the RGB image is 0 to 1. After another Conv1 convolutional layer, the size becomes 32×512×512. The downsampling layer in the neural network encoding process doubles the first dimension and halves the last two dimensions. Residual convolution (ResConv) performs feature transformation without changing the data shape. After three downsampling-ResConv layers, the data size becomes 256×64×64. The importance-aware layer generates 256 weights and multiplies them by 256 64×64 feature maps, redistributing the importance of the features without changing the data size. Then, the compression layer compresses the data size to γ×32×32 according to a pre-set compression ratio γ. The feature map is then flattened into a one-dimensional bitstream of length 4096γ and sent to the FSO channel for transmission.

[0055] After receiving the signal, the one-dimensional bitstream is rearranged into a γ×32×32 feature map. The upsampling in the neural network decoding process halves the first dimension and doubles the second and third dimensions. The rearranged features first pass through a decompression layer, generating a size of 256×64×64. Then, after three upsampling steps (ResConv), the data size becomes 32×512×512. Finally, after Conv2 convolution and denormalization layers, the hyperspectral data is restored to a size of 31×512×512, with values ​​ranging from 0 to 65536. Dense connection processing is used in both the neural network encoding and decoding processes to improve the efficiency of the network encoding.

[0056] The design and generation of DOE, such as Figure 6 As shown, firstly, a 1024×1024 floating-point matrix is ​​generated, with each floating-point number ranging from 0 to 1. Then, regions greater than 512 units from the center are set to zero, and all parameters less than or equal to 512 units from the center are set as trainable parameters, thus generating a circular DOE structure with trainable parameters. Next, the floating-point sequence is magnified 16 times and quantized using rounding, with all values ​​quantized to integers. The quantized values ​​are then converted into the material height of the DOE, and a height map is generated based on this material height for actual DOE production.

[0057] In the end-to-end joint training process, the network first initializes a 1024×1024 trainable parameter map. Based on this parameter map, a physical height map of the Design of Effect (DOE) is generated, and then the point spread function (PSF) at different wavelengths is calculated. The hyperspectral data cube is convolved with and superimposed with the PSF of each wavelength to generate a two-dimensional coded image. The generated two-dimensional coded image is then processed through a neural network encoding-channel transmission-neural network decoding process to reconstruct the hyperspectral image. The recovered hyperspectral image and the original hyperspectral image are used to calculate the mean squared error (MSE) loss function. The entire process is completely differentiable; therefore, the gradient of the loss function of the reconstructed network can be backpropagated to the 1024×1024 parameter map, directly optimizing and updating the DOE parameters through gradient descent. During backpropagation, the gradient of the loss function is backpropagated sequentially through the decoding network, the channel model, and the encoding network, updating the weights of each layer of the neural network. To enhance the system's robustness to FSO channel noise, the parameter update of the AT-IHAT system is divided into three stages, see [link to relevant documentation]. Figure 7 and Figure 8 The system operates in three phases: Phase 1 is the parameter pre-training phase, where DOE parameters and network parameters are trained and updated without adding FSO noise, enabling the system to achieve optimal hyperspectral information extraction and compression capabilities. Phase 2 is the network-channel interaction phase, where the DOE parameters trained in Phase 1 are fixed. During training, FSO channel noise is added, and the network parameters are updated, allowing the system to initially achieve robustness against FSO channel noise. Phase 3 is the DOE-channel interaction phase, where the encoding / decoding network parameters obtained in Phase 2 are fixed, and the DOE parameters are updated only when FSO channel noise is added. Figure 9 This is a block diagram of the direct-modulation, direct-detection (DSO) FSO communication system used. The one-dimensional bitstream output by the neural network generates an analog driving electrical signal using an arbitrary waveform generator. This signal controls a Mach-Zehnder modulator to modulate a 1550nm laser beam generated by the laser, thereby loading data onto the optical signal. The generated optical signal is amplified by an erbium-doped fiber amplifier and then transmitted to the receiver in the FSO channel. At the receiver, the optical signal is converted into an electrical signal by a photodetector, sampled by a DSO, and the received data is obtained. The received data is then input into a decoding network to reconstruct the hyperspectral image data.

[0058] A comparison of the transmission performance of the AT-IHAT scheme and traditional schemes in the FSO channel, such as... Figure 10 As shown, all schemes have the same transmission rate. Two sub-figures illustrate the transmission performance comparison under different turbulence intensities; the horizontal axis represents the number of transmissions, with a total of 40 transmissions. The evaluation metric is Structural Similarity Index Measure (SSIM). The traditional scheme being compared uses JPEG2000 as the source code and LDPC as the channel code. Figure 10It can be observed that when the turbulence intensity is weak, both the proposed scheme and the traditional scheme can achieve stable hyperspectral image restoration, but the proposed scheme has higher restoration quality. As the turbulence intensity increases, the restoration performance of the traditional scheme becomes unstable, and spectral image restoration fails in some cases. In contrast, the proposed scheme maintains stable hyperspectral information restoration across multiple transmissions, demonstrating strong resistance to turbulence changes.

[0059] This application also provides an electronic device, which may include a processor, a memory, a receiver, and a transmitter. The processor is used to execute the hyperspectral information acquisition and transmission method against atmospheric turbulence mentioned in the above embodiments. The processor and the memory can be connected via a bus or other means, taking a bus connection as an example. The receiver can be connected to the processor and the memory via wired or wireless means.

[0060] The processor can be a central processing unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0061] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the anti-atmospheric turbulence hyperspectral information acquisition and transmission method described in the embodiments of this application. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions, and modules stored in the memory, thereby realizing the anti-atmospheric turbulence hyperspectral information acquisition and transmission method described in the above method embodiments.

[0062] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0063] The one or more modules are stored in the memory, and when executed by the processor, the hyperspectral information acquisition and transmission method against atmospheric turbulence described in the embodiment is executed.

[0064] In some embodiments of this application, the user equipment may include a processor, a memory, and a transceiver unit. The transceiver unit may include a receiver and a transmitter. The processor, memory, receiver, and transmitter may be connected via a bus system. The memory is used to store computer instructions, and the processor is used to execute the computer instructions stored in the memory to control the transceiver unit to send and receive signals.

[0065] As one implementation method, the functions of the receiver and transmitter in this application can be implemented by transceiver circuits or dedicated transceiver chips, and the processor can be implemented by dedicated processing chips, processing circuits or general-purpose chips.

[0066] As another implementation approach, the server provided in this application embodiment can be implemented using a general-purpose computer. That is, the program code implementing the processor, receiver, and transmitter functions is stored in memory, and the general-purpose processor implements the processor, receiver, and transmitter functions by executing the code in memory.

[0067] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned hyperspectral information acquisition and transmission method against atmospheric turbulence. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0068] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the aforementioned method for acquiring and transmitting hyperspectral information against atmospheric turbulence.

[0069] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. The programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave.

[0070] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0071] In this application, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0072] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to the embodiments of this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for acquiring and transmitting hyperspectral information resistant to atmospheric turbulence, characterized in that, The method includes: The hyperspectral image is encoded and compressed using a preset diffractive optical element to obtain a three-channel RGB image corresponding to the hyperspectral image; The three-channel RGB image is input into a preset encoding neural network, which performs normalization, convolution, downsampling, dense connection, importance-aware weighting, and compression on the three-channel RGB image, and outputs a one-dimensional feature stream. Based on a preset FSO channel, the one-dimensional feature stream is transmitted to the receiving end, so that the receiving end rearranges the received one-dimensional feature stream into a feature map of a preset shape, and inputs the feature map of the preset shape into a preset decoding neural network, and the decoding neural network performs decompression, upsampling, dense connection, convolution and denormalization processing on the feature map of the preset shape, and outputs the reconstructed hyperspectral image corresponding to the hyperspectral image; wherein, the diffractive optical element, the encoding neural network and the decoding neural network are pre-trained in a channel model containing atmospheric turbulence noise based on an end-to-end joint training method.

2. The method according to claim 1, characterized in that, The encoded neural network includes: The normalization layer is used to normalize the three-channel RGB image to obtain the corresponding normalized three-channel RGB image; The first convolutional layer is used to increase the number of channels of the normalized three-channel RGB image to a preset first number of channels to obtain the first convolutional feature map; Multiple cascaded downsampling layers are used to downsample the first convolutional feature map to obtain a downsampled feature map with a preset number of second channels; wherein, the downsampling operation includes halving the spatial size of the feature map and doubling the number of channels; The first dense connection layer is used to copy the feature maps before entering each of the downsampling layers and convolve them based on a plurality of branch convolutional layers of different preset lengths to obtain branch feature maps corresponding to each of the branch convolutional layers. The branch feature maps are then added to the downsampling feature maps to obtain the first dense connection feature map. The first dense connection layer includes the same number of branch convolutional layers as the downsampling layers. The importance-aware layer is used to generate corresponding weights for the first densely connected feature maps of each channel, and perform channel-by-channel weighting to obtain the importance-aware feature maps. The compression layer is used to compress the size of the importance-aware feature map to the target dimension according to a preset compression ratio, and flatten the compressed importance-aware feature map to output the one-dimensional feature code stream.

3. The method according to claim 1, characterized in that, The decoding neural network includes: The decompression layer is used to decompress the received feature map of the preset shape to obtain the decompressed feature map; Multiple cascaded upsampling layers are used to upsample the decompressed feature map to obtain an upsampled feature map; wherein, the upsampling operation includes doubling the spatial size of the feature map and halving the number of channels to restore the spatial resolution of the feature map; The second dense connection layer is used to copy the feature maps before entering each of the upsampling layers and convolve them based on multiple pre-set transposed convolutional layers of different lengths to obtain transposed convolutional feature maps corresponding to each of the transposed convolutional layers. The transposed convolutional feature maps are then added to the upsampling feature maps to obtain the second dense connection feature map. The second dense connection layer includes the same number of transposed convolutional layers as the upsampling layers. The second convolutional layer is used to adjust the number of channels of the second densely connected feature map to a preset third number of channels to obtain the second convolutional feature map; The denormalization layer is used to denormalize the second convolutional feature map and output the reconstructed hyperspectral image.

4. The method according to claim 1, characterized in that, The end-to-end joint training specifically includes: The first update step is performed as follows: a physical height map of the diffractive optical element is generated based on the current trainable parameter map; the point spread function at different wavelengths is calculated based on the physical height map; the original hyperspectral image is convolved with and superimposed with the point spread functions of each wavelength to generate a two-dimensional coded image; the two-dimensional coded image is sequentially encoded by a coding neural network, transmitted through a channel model containing atmospheric turbulence noise, and decoded by a decoding neural network to obtain the recovered hyperspectral image; the mean square error between the recovered hyperspectral image and the original hyperspectral image is calculated as the loss function. The second update step is performed: the gradient of the loss function is propagated back to the trainable parameter map of the diffractive optical element by passing the decoding neural network, the channel model, and the encoding neural network in sequence through the gradient descent algorithm, so as to synchronously update the parameters of the diffractive optical element, the weights of the encoding neural network, and the weights of the decoding neural network. Determine whether the loss function has converged; if it has not converged, continue to execute the first update step and the second update step; until the loss function converges, use the updated diffractive optical element, encoding neural network and decoding neural network as the diffractive optical element, encoding neural network and decoding neural network for hyperspectral information processing.

5. The method as described in claim 4, characterized in that, The end-to-end joint training also includes: Perform the neural network update step: fix the parameters of the synchronously updated diffractive optical element, and update the weights of the decoding neural network based on the added atmospheric turbulence noise and the loss function. Determine whether the loss function has converged; if it has not converged, continue to execute the first update step and the neural network update step; until the loss function converges, and use the updated encoding neural network and decoding neural network as the trained encoding neural network and decoding neural network; Perform the diffractive optical element update step: fix the weights of the decoding neural network obtained from the neural network update step, and update the parameters of the diffractive optical element based on the added atmospheric turbulence noise and the loss function; Determine whether the loss function has converged; if it has not converged, continue to execute the first update step and the diffractive optical element update step; until the loss function converges, and use the updated diffractive optical element as the trained diffractive optical element.

6. The method according to claim 1, characterized in that, The hyperspectral image includes information from different spectral bands; Correspondingly, the step of encoding and compressing the hyperspectral image using a preset diffraction optical element to obtain a three-channel RGB image corresponding to the hyperspectral data cube includes: The information of different bands in the hyperspectral image is encoded using a preset diffraction optical element to generate a point spread function corresponding to the information of each band. Based on each of the point spread functions, the hyperspectral image is compressed using a preset RGB camera to obtain the three-channel RGB image.

7. The method according to claim 1, characterized in that, The transmission of the one-dimensional feature code stream to the receiving end based on the preset FSO channel includes: The one-dimensional feature code stream is modulated onto an optical carrier to generate an optical signal carrying the one-dimensional feature code stream information; and the optical signal is transmitted to the receiving end through the FSO optical channel so that the receiving end can perform photoelectric conversion and sampling on the received optical signal to obtain the recovered one-dimensional feature code stream.

8. An electronic device, characterized in that, It includes a processor and a memory; when the processor executes the running program stored in the memory, it implements the hyperspectral information acquisition and transmission method against atmospheric turbulence as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the hyperspectral information acquisition and transmission method against atmospheric turbulence as described in any one of claims 1 to 7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the hyperspectral information acquisition and transmission method against atmospheric turbulence as described in any one of claims 1 to 7.

Images