An AI-based on-board remote sensing image compression method and system

Abstract

The application discloses an AI-based satellite-borne remote sensing image compression method and system, and relates to the technical field of image compression, which comprises the following steps: acquiring an original remote sensing image, decoupling a structure tensor and a spectral curve tensor; encoding the spectral curve tensor into a diagnostic physical feature vector, searching for a match in a spectral prototype knowledge base, outputting an anchor index and a typical spectral curve, and calculating a spectral decoupling deviation parameter between the spectral curve tensor and the typical spectral curve; extracting a spatial texture feature to generate a gating weight vector, and encoding to generate a structure code stream; performing adaptive quantization and entropy encoding on the spectral decoupling deviation parameter to generate a spectral code stream; integrating the structure code stream, the anchor index and the spectral code stream into a compressed code stream; reconstructing the structure tensor according to the compressed code stream, reconstructing the spectral curve tensor according to the spectral code stream, and reconstructing the remote sensing image. The application aims to solve the problem that the existing AI compression method destroys the spectral fidelity of a remote sensing image, resulting in the loss of quantitative analysis value.

Description

Technical Field

[0001] This invention relates to the field of image compression technology, and in particular to an AI-based method and system for compressing spaceborne remote sensing images. Background Technology

[0002] Spaceborne remote sensing imagery is costly to acquire and generates large amounts of data, necessitating efficient compression for transmission. End-to-end image compression methods based on deep learning, such as convolutional autoencoders, have attracted significant attention due to their rate-distortion performance surpassing traditional coding standards. However, the optimization objective of these methods is pixel-level mathematical distortion (such as mean square error) or human-perceived quality, which fundamentally deviates from the core value of remote sensing imagery (quantitative applications).

[0003] Specifically, mean squared error optimization introduces unpredictable spectral distortions, while methods such as generative adversarial networks (GANs) can easily produce realistic but false "texture illusions," severely damaging the true spectral characteristics of ground features. This results in compressed images that, while visually appealing, cannot support high-value quantitative analyses such as ground feature classification, vegetation index inversion, and change detection, rendering the data unusable for scientific applications. Furthermore, to adapt to satellite deployment, mainstream lightweight and efficient entropy estimation techniques are primarily designed for general-purpose images, lacking sufficient ability to model non-Gaussian and heavy-tailed distributions, further exacerbating information loss under complex remote sensing textures.

[0004] Therefore, there is an urgent need for a spaceborne AI compression method that aims to maintain high compression ratio and low computational complexity while ensuring the physical properties and quantitative analysis value of compressed images in principle. Summary of the Invention

[0005] This invention provides an AI-based method for compressing spaceborne remote sensing images, comprising: Acquire raw remote sensing images and perform structure-spectral decoupling to generate structure tensors and spectral curve tensors; The spectral curve tensor is encoded as a diagnostic physical feature vector, a match is retrieved in the spectral prototype knowledge base, the anchor index and typical spectral curve are output, and the spectral decoupling deviation parameter between the spectral curve tensor and the typical spectral curve is calculated. Extract spatial texture features from the structure tensor to generate a gated weight vector, and encode it to generate a structured bitstream; Using the gated weight vector, adaptive quantization and entropy coding are performed on the spectral decoupling deviation parameter to generate a spectral bitstream; the structure bitstream, anchor index, and spectral bitstream are then combined into a compressed bitstream. The ground-based decoding unit reconstructs the structure tensor based on the compressed bitstream, and reconstructs the spectral curve tensor based on the anchor point index and spectral bitstream, and then synthesizes and reconstructs the remote sensing image.

[0006] The aforementioned AI-based spaceborne remote sensing image compression method encodes spectral curve tensors into diagnostic physical feature vectors, retrieves matches in a spectral prototype knowledge base, outputs anchor point indices and typical spectral curves, and calculates spectral decoupling deviation parameters between the spectral curve tensor and the typical spectral curves, including: A spectral prototype knowledge base is constructed based on standard ground object spectral data, and after the parameters are solidified, it is simultaneously deployed to the spaceborne encoding end and the ground decoding end. Diagnostic physical feature vectors of the original spectral curves of each pixel are extracted from the spectral curve tensor. Land cover category matching is performed in the coarse-grained layer of the spectral prototype knowledge base, and the coarse-grained prototype index is output. In the fine-grained prototype set corresponding to the matching coarse-grained node, a fine spectral shape retrieval is performed, the matching prototype parameters are output, the matching confidence is evaluated, and the encoding path is determined. For pixels with acceptable matching confidence, the spectral decoupling deviation parameter is calculated based on the original spectrum and the matching prototype spectrum.

[0007] The aforementioned AI-based spaceborne remote sensing image compression method constructs a spectral prototype knowledge base based on standard ground feature spectral data, and after solidifying parameters, synchronously deploys it to both the spaceborne encoding end and the ground decoding end, including: Collect and preprocess multi-source standard ground cover spectral samples, and calculate the basic statistical parameters of each ground cover category; A two-level tree-structured spectral prototype knowledge base is constructed based on physical feature vectors.

[0008] The aforementioned AI-based spaceborne remote sensing image compression method extracts diagnostic physical feature vectors from the original spectral curves of each pixel based on the spectral curve tensor, performs land cover category matching in a coarse-grained layer of the spectral prototype knowledge base, and outputs a coarse-grained prototype index, including: The original spectral curve vector of a single pixel is input into the physical feature encoder, and the corresponding diagnostic physical feature vector is output. Calculate the similarity between the query features and the coarse-grained prototype nodes, select the optimal node, and output the coarse-grained prototype index.

[0009] The aforementioned AI-based spaceborne remote sensing image compression method performs fine-grained spectral shape retrieval in the fine-grained prototype set corresponding to matching coarse-grained nodes, outputs matching prototype parameters, evaluates matching confidence, and determines the encoding path, including: Within the fine-grained candidate set of coarse-grained matching nodes, the fine-grained prototype node with the smallest spectral angular distance is selected as the best matching result, and the index and storage parameters of the corresponding node are output. Calculate the matching confidence score and compare it with a preset threshold. Based on the comparison results, divide the regular coding path into the adaptive backtracking coding path.

[0010] The aforementioned AI-based spaceborne remote sensing image compression method extracts spatial texture features of the structure tensor to generate a gated weight vector and encodes it to generate a structured bitstream, including: The structural tensor is input into a lightweight 2D convolutional structured flow encoder to extract spatial texture feature maps. The spatial texture feature map is input into the gated generation network to generate a gate weight vector used to adjust the spectral coding accuracy; Quantization and entropy encoding are performed on the spatial texture feature map to generate a structured bitstream.

[0011] The aforementioned AI-based spaceborne remote sensing image compression method inputs a spatial texture feature map into a gated generation network to generate a gated weight vector for adjusting the spectral coding accuracy, including: A gated generative network is used to perform channel transformation and global pooling on the spatial texture feature map, and outputs a gated weight vector. A quantization step size adaptive modulation mechanism is established based on the gated weight vector to control the encoding precision driven by spatial texture.

[0012] The aforementioned AI-based spaceborne remote sensing image compression method performs quantization and entropy encoding on spatial texture feature maps to generate a structured bitstream, including: Perform element-wise uniform scalar quantization on the spatial texture feature map to obtain a discretized quantized feature map; A hyper-prior entropy model is used to perform probability estimation and arithmetic coding on the quantized feature map to generate a structured bitstream.

[0013] The aforementioned AI-based spaceborne remote sensing image compression method utilizes a gated weight vector to perform adaptive quantization and entropy coding on the spectral decoupling deviation parameters to generate a spectral bitstream; it then synthesizes the structure bitstream, anchor index, and spectral bitstream into a compressed bitstream, including: For the spectral decoupling deviation parameters of regular path pixels, adaptive step-size quantization is performed using a gated weight vector; Using the matched prototype spectral curve as a condition, conditional entropy encoding is performed on the quantized spectral decoupling deviation parameter to generate a regular path spectral bitstream; Quantization and entropy coding are performed on the compact feature vectors of the backtrack path pixels to generate the backtrack path spectral bitstream; The structured bitstream, prototype index data, regular path spectral bitstream, and fallback path spectral bitstream are combined to generate a complete compressed bitstream.

[0014] An AI-based spaceborne remote sensing image compression system includes: The structure-spectral decoupling module is used to acquire raw remote sensing images and perform structure-spectral decoupling to generate structure tensors and spectral curve tensors. The spectral prototype retrieval module is used to encode the spectral curve tensor into a diagnostic physical feature vector, retrieve a match in the spectral prototype knowledge base, output the anchor index and typical spectral curve, and calculate the spectral decoupling deviation parameter between the spectral curve tensor and the typical spectral curve. The gating precision allocation module is used to extract the spatial texture features of the structure tensor to generate the gating weight vector and encode it to generate the structure bitstream; The adaptive compression coding module is used to perform adaptive quantization and entropy coding on the spectral decoupling deviation parameters using the gated weight vector to generate a spectral bitstream; and to combine the structure bitstream, anchor index and spectral bitstream into a compressed bitstream. The bitstream decoding and reconstruction module is used by the ground decoding end to reconstruct the structure tensor based on the compressed bitstream, reconstruct the spectral curve tensor based on the anchor point index and the spectral bitstream, and synthesize and reconstruct the remote sensing image.

[0015] The beneficial effects achieved by this invention are as follows: By constructing a spectral prototype knowledge base and using a coding paradigm based on reference comparison, the information to be encoded is transformed from a complete spectrum into deviation parameters from typical land cover prototypes, thus fundamentally avoiding damage to the spectral physical properties. By decoupling the deviation parameters through layered spectral analysis, spectral differences are decomposed into three-level components: amplitude, shape, and diagnostics. Priority is given to protecting the spectral features that are key to quantitative inversion, ensuring that the compressed image can still support high-value quantitative analysis tasks such as land cover classification and vegetation index calculation, thus achieving a balance between compression efficiency and quantitative fidelity. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart of an AI-based spaceborne remote sensing image compression method provided in Embodiment 1 of this application.

[0018] Figure 2 This is a schematic diagram of an AI-based spaceborne remote sensing image compression system provided in Embodiment 2 of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1 like Figure 1 As shown, Embodiment 1 of this application provides an AI-based spaceborne remote sensing image compression method, including: S1: Acquire the original remote sensing image and perform structure-spectral decoupling to generate the structure tensor and spectral curve tensor; The system acquires raw remote sensing images collected by spaceborne sensors. Where H and W represent the spatial height and width of the image, respectively, and C represents the number of spectral bands. The spatial height and width correspond to the spatial resolution of the image, and the number of spectral bands corresponds to the number of imaging channels of the sensor. The pixel values ​​of the original remote sensing image are the radiance values ​​of the corresponding ground features in the corresponding spectral band, or the surface reflectance values ​​after radiometric calibration and atmospheric correction.

[0021] Next, the mean value is calculated along the spectral dimension of the entire band of the original remote sensing image to generate the structure tensor. This structure tensor is used to characterize the spatial texture information of an image. It captures the common variation trend of reflectivity of ground features across the entire spectral range and is a compact representation of spatial structural features such as edges, shapes, and textures.

[0022] Subsequently, at each pixel location, the system subtracts the corresponding structure tensor value from the spectral values ​​of each band in the original remote sensing image to generate the spectral curve tensor. This is used to characterize the pure spectral shape features of ground features after removing the influence of overall intensity. In the spectral curve tensor, each spatial location corresponds to a set of numerical sequences of equal length to the number of bands, which constitute the spectral curve of that pixel. Its values ​​only reflect the spectral response morphology of the ground feature and are independent of the overall brightness level of the pixel.

[0023] Through the above structure-spectral decoupling process, the original remote sensing image is decomposed into two data streams with independent physical meanings: spatial structure and spectral morphology. The two data streams correspond to the spatial geometric attributes and spectral attributes of the ground objects, respectively, thus eliminating the coupling redundancy between spatial intensity and spectral shape.

[0024] S2: Encode the spectral curve tensor into a diagnostic physical feature vector, retrieve a match in the spectral prototype knowledge base, output the anchor index and typical spectral curve, and calculate the spectral decoupling deviation parameter between the spectral curve tensor and the typical spectral curve. The process involves encoding the spectral curve tensor into a diagnostic physical feature vector, retrieving a match from the spectral prototype knowledge base, outputting the anchor index and a typical spectral curve, and calculating the spectral decoupling deviation parameter between the spectral curve tensor and the typical spectral curve. This includes the following sub-steps: S21: Construct a spectral prototype knowledge base based on standard ground object spectral data, solidify the parameters, and then deploy it synchronously to the spaceborne encoding end and the ground decoding end; The process involves constructing a spectral prototype knowledge base based on standard ground feature spectral data, solidifying parameters, and then simultaneously deploying it to both the spaceborne encoding end and the ground decoding end. This includes the following sub-steps: S211: Collect and preprocess multi-source standard ground cover spectral samples, and calculate the basic statistical parameters of each ground cover category; We collected standard ground cover spectral data covering typical terrestrial ecosystems worldwide. Data sources included public spectral databases and historical observation spectral datasets of target sensors. The samples covered typical ground cover categories and their subcategories, such as vegetation, soil, water, snow and ice, artificial features, shadows, clouds, and bare rocks.

[0025] Preprocessing operations such as radiometric normalization, bad band removal, and noise smoothing are performed on all samples to obtain a clean sample set.

[0026] For each land cover category, the prototype spectral curve of the category is obtained by calculating the band-by-band arithmetic mean, and the spectral covariance matrix of the category is calculated by the unbiased sample statistical method. This matrix is ​​used to characterize the natural variation range of the spectrum within the category and the correlation between bands.

[0027] The prototype spectral curves of each category are input into the physical feature encoder to extract the corresponding diagnostic physical feature vectors, and the covariance matrix of each category of physical features is calculated simultaneously.

[0028] The physical feature encoder is a one-dimensional convolutional neural network used to map the input C-dimensional spectral curve into a D-dimensional diagnostic physical feature vector. Specifically, the encoder adopts a cascaded structure of three one-dimensional convolutional layers and two fully connected layers; the kernel sizes of the three convolutional layers are 7, 5, and 3 respectively, the stride of each convolution is 1, and the padding method is equal-length padding. After each convolution, a LeakyReLU activation function and a layer normalization module are applied; the output of the convolutional layers is global average pooled and then applied to two fully connected layers with dimensions of 32 and D respectively, finally outputting a D-dimensional diagnostic physical feature vector.

[0029] In this embodiment This corresponds to key quantitative remote sensing diagnostic features such as the wavelength position of the chlorophyll absorption peak, chlorophyll absorption depth, red edge position, red edge slope, water absorption depth, and cellulose absorption depth.

[0030] The encoder receives a single C-dimensional spectral curve vector as input, extracts local shape features and structured features such as absorption valleys and reflection peaks from the spectral curve through a convolutional layer, and then maps them through a fully connected layer into a low-dimensional feature vector that corresponds one-to-one with the physical diagnostic features. Each dimension of the output vector corresponds to a specific physical diagnostic feature quantization value.

[0031] The encoder takes spectral samples from the standard ground cover spectral database as input, uses the measured values ​​of the physical diagnostic features corresponding to the samples as supervision signals, and performs supervised pre-training using the mean square error loss function; after training, the network parameters are fixed.

[0032] S212: Construct a two-level tree-structured spectral prototype knowledge base based on physical feature vectors.

[0033] Using the prototype physical feature vectors of various object subclasses as clustering objects, a hierarchical clustering algorithm is used to construct a tree-like spectral prototype knowledge base (SPKB). The knowledge base is divided into a two-level structure of coarse-grained layer and fine-grained layer.

[0034] At the fine-grained level, each land cover subclass that has completed statistics corresponds to a fine-grained prototype node, and each fine-grained node k stores the fine-grained prototype spectral curve. Fine-grained spectral covariance matrix With fine-grained prototype physical feature vector ; The above three types of parameters are all directly derived from the statistical and feature extraction results of the corresponding land cover subclasses, providing a benchmark for fine-grained spectral retrieval and deviation parameter calculation.

[0035] In the coarse-grained layer, fine-grained nodes are merged layer by layer from bottom to top, using the inter-class Mahalanobis distance as the merging criterion, until the preset number of coarse-grained categories is reached. In this embodiment It corresponds to eight major categories of land features: vegetation, soil, water, ice and snow, artificial features, shadows, clouds, and bare rocks.

[0036] For each coarse-grained node j, the corresponding parameters of all its subordinate fine-grained nodes are summarized, and four types of benchmark parameters are obtained through aggregated statistical calculation: coarse-grained physical feature vector. Covariance matrix of coarse-grained physical features , coarse-grained prototype spectral curves and coarse-grained spectral covariance matrix; The vector parameters are obtained by taking the arithmetic mean of the corresponding parameters of all subordinate fine-grained nodes, and the covariance matrix is ​​obtained by statistically analyzing the unbiased sample covariance of the corresponding parameters of all subordinate fine-grained nodes. These four types of parameters together serve as the physical anchoring benchmark for coarse-grained retrieval.

[0037] Subsequently, the initialized spectral prototype knowledge base is used as a fine-tunable parameterized module and spliced ​​with the encoding and decoding modules of the compressed network. End-to-end joint training is performed with the joint optimization objectives of spectral reconstruction accuracy and quantitative remote sensing task fidelity.

[0038] During training, all parameters of the coarse-grained layer and the physical feature vector of the fine-grained layer remain fixed as stable physical anchoring references; only the prototype spectral curve and spectral covariance matrix of the fine-grained layer are fine-tuned with a low learning rate.

[0039] If the training learning rate of the compressed network backbone is The learning rate of the knowledge base spectral parameters Set as ;in, The scaling factor for the knowledge base learning rate is 0.01 in this embodiment. Its function is to constrain the knowledge base to only make small adjustments near the initial physical prior, so as to avoid overfitting to specific deviations in the training data, while adapting to the spectral response characteristics of the target sensor and typical atmospheric conditions.

[0040] After joint fine-tuning, all parameters of the spectral prototype knowledge base (including the tree-structured hierarchical index, physical feature vectors of each node, prototype spectral curves, and spectral covariance matrix) are solidified into parameter files and deployed to the non-volatile storage units of the spaceborne encoding end and the ground decoding end, respectively. The spaceborne end and the ground end hold completely identical copies of the knowledge base, ensuring that the prototype data used in the online retrieval stage and the decoding and reconstruction stage are completely consistent, avoiding reconstruction errors caused by inconsistencies between the two ends.

[0041] During system operation, the ground system performs cluster analysis and validity verification on pixels marked as uncovered types in the observation mission. Newly verified land cover categories are inserted as new fine-grained nodes under the corresponding coarse-grained nodes, updating the knowledge base parameters. The updated knowledge base is incrementally synchronized to the satellite-borne terminal via the space-ground uplink, enabling dynamic iterative optimization of the knowledge base.

[0042] S22: Extract the diagnostic physical feature vector of the original spectral curve of each pixel based on the spectral curve tensor, perform land cover category matching in the coarse-grained layer of the spectral prototype knowledge base, and output the coarse-grained prototype index. The process includes extracting diagnostic physical feature vectors from the original spectral curves of each pixel based on the spectral curve tensor, performing land cover category matching in the coarse-grained layer of the spectral prototype knowledge base, and outputting a coarse-grained prototype index. This process comprises the following sub-steps: S221: Input the original spectral curve vector of a single pixel into the physical feature encoder and output the corresponding diagnostic physical feature vector; From the spectral curve tensor Extracting the original spectral curve vector pixel by pixel Where H and W represent the spatial height and width of the image, respectively, and C represents the total number of spectral bands. Specifically, all spatial locations are traversed in row-major order. For each spatial coordinate (x, y), the values ​​of all C bands at that location are extracted along the spectral dimension, and a one-dimensional original spectral curve vector is formed in band order.

[0043] Will Input the pre-trained physical feature encoder Output D-dimensional diagnostic physical feature vector .

[0044] The network structure, parameter configuration, and pre-training method of this encoder are the same as those of the physical feature encoder of S211.

[0045] S222: Calculate the similarity between the query feature and the coarse-grained prototype node, select the optimal node, and output the coarse-grained prototype index.

[0046] Access the coarse-grained hierarchy of the spectral prototype knowledge base, which contains Each coarse-grained prototype node. Stores coarse-grained physical feature vectors Covariance matrix of coarse-grained physical features Using Mahalanobis distance as the similarity metric, the similarity is calculated. The distance values ​​between the physical feature vectors of each coarse-grained prototype node and the target node are recorded. The coarse-grained prototype node with the smallest Mahalanobis distance is selected as the matching result, and its coarse-grained index is recorded. And the subsequent search scope will be limited to the set of fine-grained child nodes under this node.

[0047] S23: Perform fine spectral shape retrieval in the set of fine-grained prototypes corresponding to the matching coarse-grained nodes, output the matching prototype parameters and evaluate the matching confidence to determine the encoding path; The process involves performing a fine-grained spectral shape retrieval within the set of fine-grained prototypes corresponding to the coarse-grained nodes, outputting the matching prototype parameters, evaluating the matching confidence, and determining the encoding path. This includes the following sub-steps: S231: Within the fine-grained candidate set of coarse-grained matching nodes, select the fine-grained prototype node with the smallest spectral angular distance as the best matching result, and output the index and storage parameters of the corresponding node. Based on coarse-grained index In the SPKB tree structure, locate all fine-grained prototype nodes belonging to the corresponding coarse-grained node to form a fine-grained candidate set, with the number of candidate nodes being [number missing]. This is much smaller than the total number of nodes in the knowledge base.

[0048] Each fine-grained prototype node k stores the prototype spectral curve. The curve is initialized by the band-wise mean of the standard spectral samples of the corresponding ground object subclass in S211, and fine-tuned with a low learning rate during the end-to-end joint training phase to adapt to the spectral response characteristics of the target sensor.

[0049] Using spectral angular distance as a measure of spectral shape similarity, the original spectral curve is calculated. Distance values ​​from the spectral curves of each candidate prototype.

[0050] Traverse all fine-grained candidate nodes, select the node with the smallest spectral angular distance as the final matching result, and record its fine-grained index. minimum spectral angular distance Read the prototype spectral curve of the matched node from SPKB. Spectral covariance matrix Prototype physical feature vector .

[0051] Among them, the prototype spectral curve is derived from the initialization of the mean of the ground cover subclass samples and the joint fine-tuning driven by the task, and is used for the benchmark calculation of the deviation parameters. The spectral covariance matrix is ​​derived from the statistical covariance of the ground cover subclass samples, and is used for the probabilistic prior of the subsequent conditional entropy encoding. The prototype physical feature vector is calculated from the prototype spectral curve by the physical feature encoder, and remains fixed throughout the training process, and is used for confidence assessment and diagnostic feature offset calculation.

[0052] S232: Calculate the matching confidence score and compare it with a preset threshold. Based on the comparison result, divide the regular coding path into the adaptive backoff coding path.

[0053] First, calculate the mean and standard deviation of the spectral angular distances of all nodes in the current fine-grained candidate set. Combining minimum spectral angular distance Distance from diagnostic spectral data Calculate the matching confidence score The calculation formula is: in, To prevent extremely small constants with a denominator of zero, this embodiment takes the value of . ; As a diagnostic spectral distance scalar, it is obtained by summing the weighted absolute deviations of the physical characteristic vectors of the original spectrum and the prototype spectrum dimension by dimension. The calculation formula is as follows: in, To query physical feature vectors The m-th dimension component; For prototype physical feature vectors The m-th dimension component; For the first The preset weights of the diagnostic features are derived from the sensitivity coefficients determined by the quantitative assessment of the impact of feature perturbations on inversion accuracy on the independent validation set. The preset values ​​are set based on the global sensitivity analysis results of typical remote sensing inversion tasks (such as chlorophyll concentration estimation and water turbidity inversion), and can be adjusted by overload according to the specific task configuration file; D is the total dimension of the physical feature vector. when much smaller When the matching result has significant discriminative power in the candidate set, the confidence level approaches 1; otherwise, the similarity of each candidate node is close and the reliability of the matching result is low. In this context, the greater the deviation in diagnostic physical features, the worse the physical consistency of the match and the lower the confidence level.

[0054] System preset reliability threshold In this embodiment .like If the matching confidence level is deemed satisfactory, proceed to the regular encoding path and continue executing step S24; if If the current pixel spectrum is determined not to be covered by the knowledge base, it is marked as an uncovered type, and the adaptive backtracking encoding path is entered: a pre-trained lightweight spectral autoencoder is called to process the original spectral curve. Direct dimensionality reduction encoding generates compact feature vectors In this embodiment At the same time, the pixel state flag is set to flag=1, and the flag and compact feature vector are directly output to the subsequent encoding steps without performing subsequent deviation parameter calculations.

[0055] The lightweight spectral autoencoder consists of an encoder (two fully connected layers, with dimensions respectively). It consists of a ReLU activation (with a middle section) and a decoder (symmetric structure), and is pre-trained with the mean square error of spectral reconstruction as the loss function. The parameters are frozen during the joint training of the compressed network.

[0056] S24: For pixels with acceptable matching confidence, calculate the spectral decoupling deviation parameter based on the original spectrum and the matching prototype spectrum.

[0057] The optimization objective is to minimize the Euclidean distance between the original spectral curve and the prototype spectrum after amplitude scaling. A single-variable least squares optimization problem is constructed, and the optimal spectral amplitude scaling factor is obtained by solving it. The formula for calculating its closed-form solution is: in, This is the vector of the original spectral curve of the current pixel; The prototype spectral curves obtained for matching are derived from SPKB fine-grained node storage data; As a scalar, it represents the scaling factor of the overall radiative intensity of the original spectrum relative to the prototype spectrum. It corresponds to the uniform intensity change across the entire band caused by factors such as illumination conditions, terrain shadows, and atmospheric path radiation. It is the intensity component of the spectral decoupling deviation parameter.

[0058] Using spectral amplitude scaling factor Perform element-wise intensity correction on the prototype spectral curve, then perform element-wise subtraction with the original spectral curve to obtain the spectral shape residual vector. The spectral shape residual vector removes the interference of global intensity differences and only reflects the subtle spectral shape differences caused by local factors such as canopy structure, water content, and soil background. Since the prototype spectrum has captured the core spectral morphology of the subclasses of land cover, the value of this residual vector is highly concentrated near zero, exhibiting significant sparsity and low energy characteristics. It is the shape component of the spectral decoupling deviation parameter.

[0059] Diagnostic spectral distance scalar This scalar serves as a diagnostic component for the spectral decoupling deviation parameter. It quantifies the overall shift between the original and prototype spectra in key physical diagnostic features. A larger shift indicates a more significant difference between the quantitative inversion key information of the land cover and the prototype, providing a basis for subsequent adaptive quantization rate priority allocation.

[0060] Finally, the output data structures are assembled for the two types of encoded paths respectively: For regular path pixels with acceptable matching confidence, the output data structure is as follows: For backtracking path cells with unsatisfactory matching confidence, the output data structure is as follows: .

[0061] Here, flag is a pixel status flag bit, with a value of 0 indicating a normal encoding path and a value of 1 indicating a backtracking encoding path; This is a coarse-grained prototype index used to identify the node number of the land cover category to which the current pixel belongs in the coarse-grained layer of the spectral prototype knowledge base; This is a fine-grained prototype index used to identify the node number of the optimal land cover subclass matched by the current cell in the fine-grained layer of the spectral prototype knowledge base; the decoding end can use this index to look up the corresponding prototype spectral curve from the local knowledge base copy without occupying bitstream transmission bandwidth. To match the obtained fine-grained prototype spectral curve, A dimensional vector serves as the reference spectrum for calculating the spectral decoupling deviation parameter; To provide the fine-grained spectral covariance matrix for matching nodes, A dimensional matrix is ​​used to provide prior probability information about the residual distribution in the subsequent conditional entropy encoding stage; is the spectral amplitude scaling factor, which characterizes the global radiative intensity scaling factor of the original spectrum relative to the prototype spectrum, and is the intensity component of the spectral decoupling deviation parameter; Let be the spectral shape residual vector, and be The dimension vector represents the pure shape difference between the original spectrum and the prototype spectrum after removing global intensity differences, and is the shape component of the spectral decoupling deviation parameter; As a diagnostic spectral distance scalar, it quantifies the comprehensive deviation between the original spectrum and the prototype spectrum in key diagnostic physical features, and is the diagnostic component of the spectral decoupling deviation parameter; For the compact feature vector of the backtracking path, A dimensional vector is a spectral compressed representation of uncovered type pixels, used at the decoding end to reconstruct the corresponding spectral curve.

[0062] S3: Extract the spatial texture features of the structure tensor to generate a gating weight vector, and encode it to generate a structured bitstream; The structure tensor is input into the structure flow encoder to extract the spatial texture feature map. Based on the spatial texture feature map, a gating weight vector is generated to adjust the encoding accuracy of each component of the spectral decoupling deviation parameter. At the same time, quantization and entropy encoding are performed on the spatial texture feature map, and the gating weight vector and structure code stream are output.

[0063] The process of extracting spatial texture features from the structure tensor to generate a gating weight vector and encoding it to generate a structured bitstream includes the following sub-steps: S31: Input the structural tensor into a lightweight 2D convolutional structured flow encoder to extract the spatial texture feature map; Receive structure tensor The structure tensor is the mean image of the original image across all bands, capturing the common trend of reflectivity variation of ground features across all bands, and carrying spatial structure information such as the image's edges, textures, and shapes.

[0064] Structure tensor Input structure flow encoder The structured flow encoder is a lightweight two-dimensional convolutional neural network, consisting of four convolutional layers and two downsampling layers stacked alternately; the kernel size of each convolutional layer is [missing information]. The convolutional stride is 1 for all layers, and equal-length padding is used to maintain the feature map spatial size. Each convolutional layer is followed by a LeakyReLU activation function. Downsampling layers use a stride of 2. Max pooling operations are performed after the first and third convolutional layers, respectively. The number of output channels for the four convolutional layers are 32, 64, 128, and 128, respectively.

[0065] After processing by the structured flow encoder, the output spatial texture feature map is obtained. ;in, , The feature map space height and width are after two downsampling steps; This represents the total number of channels in the feature map.

[0066] Spatial texture feature map Each spatial location (u,v) corresponds to a point in the original image. The pixel block at that location encodes spatial structure information such as texture complexity, edge orientation, and contrast of the corresponding local region through its corresponding 128-dimensional feature vector. Regions with more complex textures and denser edges have larger feature vector magnitudes and more dispersed directional distributions; conversely, regions with smoother and more uniform textures have smaller feature vector magnitudes and more concentrated directional distributions. This characteristic provides a texture prior related to spatial location for the generation of subsequent gating signals.

[0067] S32: Input the spatial texture feature map into the gated generation network to generate a gated weight vector for adjusting the spectral coding accuracy; The process involves inputting the spatial texture feature map into a gated generation network to generate a gated weight vector used to adjust the spectral coding accuracy, including the following sub-steps: S321: Perform channel transformation and global pooling on the spatial texture feature map through a gated generative network, and output a gated weight vector; Spatial texture feature map Input Gated Generative Network The gated generation network consists of two cascaded pointwise convolutional layers and one global average pooling layer; the pointwise convolutional layer has a kernel size of [missing information]. Convolution operations can achieve cross-channel information fusion without changing the spatial size of the feature map.

[0068] Among them, the first layer of the gated generative network uses pointwise convolution to process the input... The channels are compressed to 64 channels, and the output is passed to the ReLU activation function for nonlinear transformation; the second layer of pointwise convolution maps the 64 channels to... Each channel outputs a value that is then fed into a Sigmoid activation function to normalize the output values ​​to the [0,1] interval. Finally, a global average pooling layer is used to calculate the arithmetic mean of the spatial dimensions of the feature map, mapping the two-dimensional spatially distributed features to a globally uniform weight vector. The specific calculation formula is as follows: in, Spatial texture feature map; The weight tensor of the first layer's pointwise convolution has a dimension of . , The bias vector is 64, representing the first-layer pointwise convolution. Both are learnable parameters of the first-layer pointwise convolution, used to perform channel-dimensional information compression and cross-channel feature fusion on the input feature map. Furthermore, due to the use of… Convolution kernel: This operation only adjusts the number of channels and does not change the spatial size of the feature map; This is the convolution operator; This is a linear rectified activation function used to perform a nonlinear transformation on the output of the first layer's pointwise convolution; The weight tensor of the second layer's pointwise convolution has dimension . , The bias vector for the second layer of pointwise convolution has dimension . These two parameters are learnable parameters of the second-layer pointwise convolution, used to map the number of feature channels to a one-to-one correspondence with the number of spectral encoded objects. Dimensions provide feature data for matching dimensions in subsequent global aggregation; This is a global average pooling operation used to calculate the arithmetic mean of the feature map output by the pointwise convolution of the second layer in the spatial dimension, compressing the two-dimensional feature map with spatial distribution attributes into a global feature vector without spatial dimension, eliminating spatial position differences, and obtaining a global control weight that can be applied to the full image spectral coding. The sigmoid activation function is used to perform an element-wise non-linear mapping on the feature vector output by global average pooling, normalizing the values ​​of all components to 0. The interval is used to ensure that the output conforms to the value constraints of the weight coefficients; For dimension The gate weight vector is used to control the quantization step size of each component of the spectral decoupling deviation parameter.

[0069] The gating generative network ultimately outputs a gating weight vector. The value range of its components is as follows: .

[0070] S322: A quantization step size adaptive modulation mechanism is established based on the gated weight vector to control the encoding precision driven by spatial texture.

[0071] Gating weight vector The dimensions correspond one-to-one with the encoded objects of the spectral decoupling deviation parameters; specifically, the former... Dimensional components Corresponding to the spectral shape residual vectors Coding precision weights for each band; Dimensional components Corresponding spectral amplitude scaling factor The coding precision weight; the first Dimensional components Corresponding diagnostic spectral distance scalar The coding precision weight.

[0072] The gated weight vector achieves adaptive adjustment of coding precision by modulating the quantization step size. Based on a preset base quantization step size... Calculate the adaptive quantization step size of the i-th encoded object after gating weight modulation. The calculation formula is: in, This is the preset base quantization step size, a global configuration parameter for the system. The gating strength scaling factor is set to 2.0 in this embodiment. It is determined by the rate-distortion performance trade-off experiment on the validation set and is used to control the adjustment of the gating weight on the quantization step size. Let i be the i-th dimension component of the gate weight vector, with a value range of 1. .when When it approaches 1, Approaching The quantization step size is taken as the reference value, and the coding accuracy reaches the preset reference level; when When it approaches 0, Approaching As the quantization step size is increased, the encoding precision is reduced accordingly, and the bit rate consumption is reduced as a result.

[0073] All parameters of the gated generation network are jointly optimized with the structured flow encoder during end-to-end training of the compression network. During training, the network automatically learns the optimal mapping relationship between texture complexity and encoding precision through backpropagation: higher gate weights are output in textured complex regions to maintain encoding precision, and lower gate weights are output in textured smooth regions to save bitrate. The spatial masking effect in textured complex regions makes compression distortion less noticeable, while spectral errors in textured uniform regions have a more significant impact on visual quality and quantitative analysis. Therefore, the network automatically tends to assign higher encoding precision to the latter.

[0074] S33: Perform quantization and entropy encoding on the spatial texture feature map to generate a structured bitstream.

[0075] The process of performing quantization and entropy encoding on the spatial texture feature map to generate a structured bitstream includes the following sub-steps: S331: Perform element-wise uniform scalar quantization on the spatial texture feature map to obtain a discretized quantized feature map; Spatial texture feature map The input quantizer performs element-wise uniform scalar quantization. Continuous input values ​​are discretized to the nearest integer index value using a fixed quantization step size. In this embodiment, the quantization step size of the structured flow is set to... The specific calculation method is as follows: in, This indicates the rounding operation; This is the quantized spatial texture feature map.

[0076] S332: The super prior entropy model is used to perform probability estimation and arithmetic coding on the quantized feature map to generate a structured bitstream.

[0077] quantized spatial texture feature map The input entropy coding module employs a super-prior entropy model for probability estimation and arithmetic coding. This super-prior entropy model consists of a super encoder and a super decoder: the super encoder extracts edge information from the quantized feature map and encodes it into an edge information bitstream; the super decoder predicts the probability distribution parameters (mean and variance) of each element in the quantized feature map based on the edge information, thereby improving the probability estimation accuracy of entropy coding and reducing bit rate consumption. Specifically, the super encoder consists of three cascaded convolutional layers, with each layer having a kernel size of [size missing]. The convolution strides are 2, 2, and 1, respectively, and the number of output channels is 128, 192, and 192, respectively. Each convolutional layer is followed by a GDN (Generalized Divide Normalization) activation function to normalize the feature distribution. GDN is a common activation function in end-to-end image compression. The system then processes the quantized spatial texture feature map... The input super encoder, after undergoing three convolutional transformations, outputs a side information feature map. This is used to characterize the global statistical distribution properties of spatial texture features. (Edge information feature map) Perform uniform scalar quantization and entropy coding to generate a side information bitstream, which is then encapsulated and transmitted along with the main structure bitstream.

[0078] The super decoder consists of three cascaded convolutional layers, with each layer having a kernel size of [size missing]. The convolution strides are 1, 2, and 2, respectively, and the number of output channels is 192, 128, and 256, respectively. Each convolutional layer is followed by an inverse GDN activation function, forming a symmetrical correspondence with the normalization operation of the super encoder. The system inputs the quantized side information feature map into the super decoder, and after layer-by-layer convolution and upsampling, it recovers the image to the same size as the super encoder. With consistent spatial dimensions, the final output is a probability distribution parameter tensor. This tensor has 256 channels, which correspond one-to-one with the 128 channels of the input features. Each feature channel corresponds to a set of mean and variance parameters, which are used to describe the Gaussian probability distribution of the feature values ​​at each spatial location under that channel.

[0079] Based on the probability distribution parameters output by the super decoder, the probability distribution corresponding to each element in the quantized feature map is determined. An arithmetic coding algorithm is used to perform adaptive arithmetic coding on the quantized spatial texture feature map to generate the main feature bitstream.

[0080] The aforementioned side information bitstream and main feature bitstream are concatenated and encapsulated according to a preset bitstream format to finally generate a complete structured bitstream. .

[0081] Gating weight vector With structured bitstream Assembled into output data, the output data structure is as follows: ;in, It is a gated weight vector used to adjust the quantization step size of each component of the spectral decoupling deviation parameter, thereby achieving adaptive coding precision allocation driven by spatial texture complexity. The structure stream, along with the spectral stream and prototype index data, is merged into a complete compressed stream and transmitted to the ground. The structure tensor is then decoded and reconstructed, and combined with the reconstructed spectral curve tensor to synthesize the final reconstructed remote sensing image.

[0082] S4: Using the gated weight vector, perform adaptive quantization and entropy coding on the spectral decoupling deviation parameters to generate a spectral bitstream; combine the structure bitstream, anchor index, and spectral bitstream into a compressed bitstream. The process involves using a gated weight vector to perform adaptive quantization and entropy coding on the spectral decoupling deviation parameters to generate a spectral bitstream. The structural bitstream, anchor index, and spectral bitstream are then combined into a compressed bitstream, including the following sub-steps: S41: For the spectral decoupling deviation parameters of regular path pixels, perform adaptive step-size quantization using a gated weight vector; Iterate through all pixels in the current image marked as a regular path (i.e., matching confidence is acceptable, flag=0). For each such pixel, extract the spectral decoupling deviation parameter, including the spectral amplitude scaling factor. Spectral shape residual vector and diagnostic spectral distance scalar .

[0083] Based on the gate weight vector Independent adaptive quantization is performed on each component of the above parameters. The correspondence between the gating weight vector and the dimension of the encoded object, as well as the calculation method of the adaptive quantization step size, have been fully explained in S322, and the results are directly quoted here.

[0084] Next, element-wise uniform scalar quantization is performed on each encoded object, discretizing the continuous values ​​into integer index values ​​through rounding. The quantized components are denoted as: quantization spectral amplitude scaling factor. Quantization spectral shape residual vector Quantitative diagnostic spectral distance scalar The aforementioned quantized integer values ​​will serve as the symbol sequence for subsequent entropy encoding.

[0085] S42: Using the matched prototype spectral curve as a condition, perform conditional entropy encoding on the quantized spectral decoupling deviation parameter to generate a regular path spectral bitstream; Construct a lightweight conditional entropy estimation network The network consists of three fully connected layers. The input dimension is C (the dimension of the prototype spectral curve), the intermediate layer dimensions are 32 and 64 respectively, and the output layer dimension is... The activation function used is ReLU. The conditional entropy estimation network is jointly trained with the encoder and decoder of the compression network backbone. During training, the prototype spectral curve is used as input, and the coding rate that minimizes the quantization deviation parameter is the optimization objective. Specifically, the distribution parameters of the network output are used to calculate the negative log-likelihood of each quantized symbol under the Laplace distribution, and the sum of these negative log-likelihoods is the code rate loss term. During training, the gradient is backpropagated to the conditional entropy estimation network via the arithmetic encoder (which uses a pass-through estimator to approximate the quantization operation) to update its fully connected layer parameters. The optimizer used is Adam, and the initial learning rate is the same as that of the main network.

[0086] Matched prototype spectral curves Input a conditional entropy estimation network, and output the probability distribution parameters of each encoded object: in, and These are the spectral shape residuals. The location and scale parameters of each component; and These are the position and scale parameters for the spectral amplitude scaling factor; and For diagnostic spectral distance scalars, the location and scale parameters are denoted as C; C represents the total number of spectral bands in the original image. This represents a conditional entropy estimation network.

[0087] The Laplace distribution is used as the probability distribution model for the quantization values ​​of each encoded object. The location and scale parameters of this distribution are determined by the prediction parameters output by the conditional entropy estimation network.

[0088] quantized symbol sequence The arithmetic coding algorithm is used to input each symbol. Based on the conditional entropy, the algorithm estimates the distribution parameters of the network output to calculate the probability value of the current symbol, performs interval subdivision and iterative coding, and generates the spectral code stream segment for that pixel. Arithmetic coding is a lossless coding method. Its basic principle is to recursively subdivide the coding interval according to the symbol probability and output the shortest binary representation. This embodiment directly adopts the standard implementation.

[0089] The prototype index set of the same cell Using fixed-length encoding, the required number of bits is Where K is the total number of nodes in the spectral prototype knowledge base. The index of the fixed-length encoding is merged with the code stream segment generated by conditional arithmetic encoding to form the complete spectral encoded data of the pixel.

[0090] The encoded data of all regular path pixels are concatenated, and synchronization markers are inserted at the pixel boundaries to form a regular path spectral code stream.

[0091] S43: Perform quantization and entropy coding on the compact feature vector of the backtrack path pixels to generate the backtrack path spectral bitstream; Iterate through all pixels in the current image marked as a backtracking path (i.e., those with unsatisfactory matching confidence in S233, flag=1). For each such pixel, extract a compact feature vector. In this embodiment .

[0092] right Perform uniform scalar quantization to obtain a compact feature vector. .

[0093] Subsequently, a factorized prior entropy model was used to analyze the quantized... Probability estimation and entropy encoding are performed. The factorized prior entropy model consists of two parts: a parameter prediction network and an arithmetic encoder.

[0094] The parameter prediction network is a three-layer fully connected cascaded structure: the first layer input is... The first layer takes a quantized feature vector index of dimension 1, outputs a 32-dimensional output, and then connects to the ReLU activation function; the second layer takes a 32-dimensional input and outputs a 64-dimensional output, which is then connected to the ReLU activation function; the third layer takes a 64-dimensional input and outputs... Dimension, where every two output values ​​correspond to the mean parameter of a feature component. and standard deviation parameter , The network outputs the probability distribution parameters of each component, assuming that each component follows an independent Gaussian distribution with a mean of [missing value]. and standard deviation Determined by the parameter prediction network.

[0095] The arithmetic encoder calculates probability values ​​based on the Gaussian distribution parameters of each component. Symbol-by-symbol arithmetic coding is performed to generate a backtracking path spectral bitstream. The factorized prior entropy model is also jointly trained with the main network. Its parameter prediction network takes the feature index as input and outputs the Gaussian distribution parameters of each component. The training objective is to minimize the coding rate of the quantized compact feature vector, which is the sum of the negative log-likelihoods of each component under the Gaussian distribution, constituting the code rate loss term. Gradients are backpropagated through a pass-through estimator, and the optimizer uses Adam.

[0096] To distinguish between the two paths, the system adds a synchronization header before each backtracked pixel bitstream, which includes a flag bit flag=1 and the spatial location index of the pixel, ensuring that the decoder can correctly distinguish and select the corresponding reconstruction method.

[0097] S44: Combine the structured bitstream, prototype index data, regular path spectral bitstream, and fallback path spectral bitstream to generate a complete compressed bitstream.

[0098] According to the preset bitstream encapsulation format, the global header information, structure bitstream, pixel type mapping table, regular path spectral bitstream, and fallback path spectral bitstream are written sequentially.

[0099] The global header information includes image spatial dimensions H and W, the number of spectral bands C, the version number of the spectral prototype knowledge base, and the length pointer of each bitstream segment; the structural bitstream carries compressed encoded data of spatial texture features; the pixel type mapping table records the encoding path flag sequence of each pixel, which the decoder uses to distinguish between regular paths and backtracking paths; the regular path spectral bitstream contains the prototype index and conditional entropy encoded data of all regular path pixels; the backtracking path spectral bitstream contains the compact feature encoded data of all backtracking path pixels.

[0100] The final compressed bitstream will be transmitted to the ground decoding end via the satellite-to-ground link.

[0101] S5: The ground decoding end reconstructs the structural tensor based on the compressed bitstream, reconstructs the spectral curve tensor based on the anchor point index and spectral bitstream, and synthesizes and reconstructs the remote sensing image. After receiving the complete compressed bitstream, the ground decoding unit parses the data units sequentially according to the preset bitstream encapsulation format; specifically, it reads the global header information and obtains the image spatial dimensions. Number of spectral bands The version number of the spectral prototype knowledge base and the length pointer of each bitstream segment are obtained; based on the length pointer, the structure bitstream, the pixel type mapping table, the regular path spectral bitstream, and the fallback path spectral bitstream are extracted respectively.

[0102] The decoding process is symmetrical between the structured bitstream input and the structured bitstream encoder. First, entropy decoding is performed on the structured bitstream to obtain the quantized spatial texture feature map. Spatial texture feature map recovered by inverse quantization Then input the structure stream decoder. The decoder is related to the structured stream encoder. It has a mirror-symmetric structure, consisting of two upsampling layers and four convolutional layers stacked alternately. After reconstruction layer by layer, the output decoded structure tensor is generated. Entropy decoding, inverse quantization, and the specific operations of the decoder are all conventional inverse processes symmetrical to those at the encoding end.

[0103] According to the cell type mapping table, the marker bits are read cell by cell. Differentiate between the normal path (flag=0) and the rollback path (flag=1), and execute the following reconstruction process for each: For regular path cells, read the prototype index. The prototype spectral curve is retrieved from a locally stored copy of the spectral prototype knowledge base. Then, conditional entropy decoding is performed on the spectral bitstream: Using the input as input, the network output probability distribution parameters are estimated through conditional entropy estimation. Based on this, arithmetic decoding is performed on the bitstream to obtain the quantization deviation parameters. Recovery via inverse quantization The system reconstructs the spectral curve of the pixel using the following formula: ; in, These are the original parameters of the encoding end. The reconstructed value after quantization and dequantization.

[0104] For backtracking path pixels: the system performs entropy decoding and inverse quantization on the spectral bitstream to obtain compact feature vectors. The decoder part of a pre-trained lightweight spectral autoencoder is input to reconstruct the spectral curve of the pixel. The structure of this autoencoder is consistent with that of the lightweight spectral autoencoder, and its decoder and encoder have a symmetrical structure.

[0105] Arrange the reconstructed spectral curves of each pixel according to their spatial positions to obtain the reconstructed spectral curve tensor. .

[0106] decode structure tensor With the reconstructed spectral curve tensor The reconstructed remote sensing image is synthesized by the reverse process of structure-spectral decoupling. The synthesized reconstructed image can be directly used for subsequent quantitative remote sensing analysis.

[0107] Example 2

[0108] like Figure 2 As shown, Embodiment 2 of this application provides an AI-based spaceborne remote sensing image compression system, including: Structure-spectral decoupling module 21: used to acquire the original remote sensing image and perform structure-spectral decoupling to generate structure tensor and spectral curve tensor; Spectral Prototype Retrieval Module 22: This module encodes the spectral curve tensor into a diagnostic physical feature vector, retrieves matches in the spectral prototype knowledge base, outputs anchor point indexes and typical spectral curves, and calculates the spectral decoupling deviation parameter between the spectral curve tensor and the typical spectral curve. Gating precision allocation module 23: used to extract the spatial texture features of the structure tensor to generate a gating weight vector and encode it to generate a structure bitstream; Adaptive compression coding module 24: Used to perform adaptive quantization and entropy coding on the spectral decoupling deviation parameters using the gated weight vector to generate a spectral bitstream; and combine the structure bitstream, anchor index and spectral bitstream into a compressed bitstream; Stream decoding and reconstruction module 25: used by the ground decoding end to reconstruct the structure tensor based on the compressed stream, reconstruct the spectral curve tensor based on the anchor point index and the spectral stream, and synthesize and reconstruct the remote sensing image; The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. An AI-based method for compressing spaceborne remote sensing images, characterized in that, include: Acquire raw remote sensing images and perform structure-spectral decoupling to generate structure tensors and spectral curve tensors; The spectral curve tensor is encoded as a diagnostic physical feature vector, a match is retrieved in the spectral prototype knowledge base, the anchor index and typical spectral curve are output, and the spectral decoupling deviation parameter between the spectral curve tensor and the typical spectral curve is calculated. Extract spatial texture features from the structure tensor to generate a gated weight vector, and encode it to generate a structured bitstream; Using the gated weight vector, adaptive quantization and entropy coding are performed on the spectral decoupling deviation parameter to generate a spectral bitstream; the structure bitstream, anchor index, and spectral bitstream are then combined into a compressed bitstream. The ground-based decoding unit reconstructs the structure tensor based on the compressed bitstream, and reconstructs the spectral curve tensor based on the anchor point index and spectral bitstream, and then synthesizes and reconstructs the remote sensing image.

2. The AI-based spaceborne remote sensing image compression method according to claim 1, characterized in that, The spectral curve tensor is encoded as a diagnostic physical feature vector. Matches are retrieved from the spectral prototype knowledge base, and anchor point indices and typical spectral curves are output. Furthermore, the spectral decoupling deviation parameter between the spectral curve tensor and the typical spectral curve is calculated, including: A spectral prototype knowledge base is constructed based on standard ground object spectral data, and after the parameters are solidified, it is simultaneously deployed to the spaceborne encoding end and the ground decoding end. Diagnostic physical feature vectors of the original spectral curves of each pixel are extracted from the spectral curve tensor. Land cover category matching is performed in the coarse-grained layer of the spectral prototype knowledge base, and the coarse-grained prototype index is output. In the fine-grained prototype set corresponding to the matching coarse-grained node, a fine spectral shape retrieval is performed, the matching prototype parameters are output, the matching confidence is evaluated, and the encoding path is determined. For pixels with acceptable matching confidence, the spectral decoupling deviation parameter is calculated based on the original spectrum and the matching prototype spectrum.

3. The AI-based spaceborne remote sensing image compression method according to claim 2, characterized in that, A spectral prototype knowledge base is constructed based on standard ground feature spectral data. After parameters are solidified, it is simultaneously deployed to both the spaceborne encoding end and the ground decoding end, including: Collect and preprocess multi-source standard ground cover spectral samples, and calculate the basic statistical parameters of each ground cover category; A two-level tree-structured spectral prototype knowledge base is constructed based on physical feature vectors.

4. The AI-based spaceborne remote sensing image compression method according to claim 2, characterized in that, Diagnostic physical feature vectors of the original spectral curves of each pixel are extracted based on the spectral curve tensor. Land cover category matching is then performed at the coarse-grained layer of the spectral prototype knowledge base, outputting a coarse-grained prototype index, including: The original spectral curve vector of a single pixel is input into the physical feature encoder, and the corresponding diagnostic physical feature vector is output. Calculate the similarity between the query features and the coarse-grained prototype nodes, select the optimal node, and output the coarse-grained prototype index.

5. The AI-based spaceborne remote sensing image compression method according to claim 2, characterized in that, A fine-grained spectral shape search is performed on the set of fine-grained prototypes corresponding to the matching coarse-grained nodes. The matching prototype parameters are output, and the matching confidence is evaluated to determine the encoding path, including: Within the fine-grained candidate set of coarse-grained matching nodes, the fine-grained prototype node with the smallest spectral angular distance is selected as the best matching result, and the index and storage parameters of the corresponding node are output. Calculate the matching confidence score and compare it with a preset threshold. Based on the comparison results, divide the regular coding path into the adaptive backtracking coding path.

6. The AI-based spaceborne remote sensing image compression method according to claim 1, characterized in that, Extracting spatial texture features from the structure tensor to generate a gating weight vector, and encoding it to generate a structured bitstream, including: The structural tensor is input into a lightweight 2D convolutional structured flow encoder to extract spatial texture feature maps. The spatial texture feature map is input into the gated generation network to generate a gate weight vector used to adjust the spectral coding accuracy; Quantization and entropy encoding are performed on the spatial texture feature map to generate a structured bitstream.

7. The AI-based spaceborne remote sensing image compression method according to claim 6, characterized in that, The spatial texture feature map is input into a gated generation network to generate a gate weight vector used to adjust the precision of spectral coding, including: A gated generative network is used to perform channel transformation and global pooling on the spatial texture feature map, and outputs a gated weight vector. A quantization step size adaptive modulation mechanism is established based on the gated weight vector to control the encoding precision driven by spatial texture.

8. The AI-based spaceborne remote sensing image compression method according to claim 6, characterized in that, Quantization and entropy encoding are performed on the spatial texture feature map to generate a structured bitstream, including: Perform element-wise uniform scalar quantization on the spatial texture feature map to obtain a discretized quantized feature map; A hyper-prior entropy model is used to perform probability estimation and arithmetic coding on the quantized feature map to generate a structured bitstream.

9. The AI-based spaceborne remote sensing image compression method according to claim 1, characterized in that, Using a gated weight vector, adaptive quantization and entropy coding are performed on the spectral decoupling deviation parameters to generate a spectral bitstream; the structure bitstream, anchor index, and spectral bitstream are then combined into a compressed bitstream, including: For the spectral decoupling deviation parameters of regular path pixels, adaptive step-size quantization is performed using a gated weight vector; Using the matched prototype spectral curve as a condition, conditional entropy encoding is performed on the quantized spectral decoupling deviation parameter to generate a regular path spectral bitstream; Quantization and entropy coding are performed on the compact feature vectors of the backtrack path pixels to generate the backtrack path spectral bitstream; The structured bitstream, prototype index data, regular path spectral bitstream, and fallback path spectral bitstream are combined to generate a complete compressed bitstream.

10. An AI-based spaceborne remote sensing image compression system, characterized in that, include: The structure-spectral decoupling module is used to acquire raw remote sensing images and perform structure-spectral decoupling to generate structure tensors and spectral curve tensors. The spectral prototype retrieval module is used to encode the spectral curve tensor into a diagnostic physical feature vector, retrieve a match in the spectral prototype knowledge base, output the anchor index and typical spectral curve, and calculate the spectral decoupling deviation parameter between the spectral curve tensor and the typical spectral curve. The gating precision allocation module is used to extract the spatial texture features of the structure tensor to generate the gating weight vector and encode it to generate the structure bitstream; The adaptive compression coding module is used to perform adaptive quantization and entropy coding on the spectral decoupling deviation parameters using the gated weight vector to generate a spectral bitstream; and to combine the structure bitstream, anchor index and spectral bitstream into a compressed bitstream. The bitstream decoding and reconstruction module is used by the ground decoding end to reconstruct the structure tensor based on the compressed bitstream, reconstruct the spectral curve tensor based on the anchor point index and the spectral bitstream, and synthesize and reconstruct the remote sensing image.

Images