Chain diffusion remote sensing hyperspectral image super-resolution system and method

Abstract

The present invention relates to the technical field of remote sensing image processing, and in particular to a chain-diffusion remote sensing hyperspectral image super-resolution system and method. In a dynamic hypergraph diffusion branch, dynamic hypergraph learning is combined with a diffusion model, and an iterative optimization process is guided by a time-varying weight space, thereby achieving joint modeling of a denoising process and feature evolution. In a differential-frequency collaborative attention branch, a differential-frequency collaborative attention mechanism is constructed, which breaks through the limitations of traditional spectral-spatial separation modeling and achieves a balance between global trend modeling and local morphology preservation. In the method provided by the present invention, a semantically constrained loss function is constructed, and a semantically driven local constraint mechanism is used to significantly improve the reconstruction quality of complex areas.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image processing, and in particular relates to a chain diffusion remote sensing hyperspectral image super-resolution system and method. Background Art

[0002] Hyperspectral remote sensing technology, with its unique ability to capture multi-band spectral information, can provide precise spectral signatures for object identification and classification. However, due to the hardware limitations of existing imaging systems, hyperspectral imagery often struggles to capture high-resolution spatial detail. This lack of spatial resolution severely limits the technology's practical application in applications such as fine-grained object identification and faint target detection.

[0003] The core scientific challenge of hyperspectral image super-resolution lies in achieving the dual optimization goals of reconstructing spatial detail while preserving spectral features. With breakthrough advances in machine learning, deep learning has demonstrated superior performance in computer vision, particularly in image quality enhancement, where it has established a new paradigm. Compared to traditional optimization methods based on physical models, deep neural networks autonomously exploit high-order nonlinear features in the data, achieving superior spatial detail reconstruction while maintaining spectral fidelity. Unlike traditional RGB images, the spectral fidelity required for hyperspectral image reconstruction significantly increases the challenge. 2D convolutions are excellent for capturing spatial information, while 3D convolutions offer unique advantages in mining spectral features. However, this specialized three-dimensional operation leads to an explosive increase in computational complexity, limiting the model's application scenarios. Thanks to their outstanding global modeling capabilities, the Transformer and its variants have demonstrated groundbreaking performance in image processing tasks. Compared to the CNN (convolutional neural network) framework, which excels at extracting local information, the Transformer, based on global self-attention, is able to establish global long-range dependencies between pixels within a single layer. Remote sensing images typically cover large observation areas, so the ability to model global information is crucial for improving reconstruction quality. In the field of hyperspectral image super-resolution, a common approach is to combine the Transformer with a 3D convolutional network to construct a learning framework that achieves the coordinated modeling of global and local information. However, the lack of explicit prior information guiding the learning process in Transformer-based frameworks significantly limits the model's learning efficiency.

[0004] Current research on remote sensing hyperspectral image super-resolution faces the following challenges:

[0005] 1) The complex spectral-spatial coupling relationships presented in remote sensing images place higher demands on the model's high-level modeling capabilities. Remote sensing images typically correspond to large observation areas and low spatial resolution, resulting in a large number of mixed pixels within the image. The spectral reflectance characteristics of mixed pixels are strongly correlated with their spatial distribution. Traditional convolutional neural networks struggle to capture global dependencies, and Transformers lack the guidance of explicit prior information.

[0006] 2) The lack of joint constraints on multi-level features often leads to blurred textures in complex areas. Effective guidance of the learning process using spectral and spatial structural characteristics at multiple scales is crucial for maintaining semantic and structural consistency. The lack of joint constraints often leads to spectral distortion and localized texture blurring. Summary of the Invention

[0007] In view of this, the present invention aims to provide a chain-diffusion remote sensing hyperspectral image super-resolution system and method, which combines dynamic hypergraph learning with the diffusion model, guides the iterative optimization process through the time-varying weight space, and realizes the joint modeling of the denoising process and feature evolution; through the differential-frequency collaborative attention mechanism, it breaks through the limitations of traditional spectral-spatial separation modeling and achieves a balance between global trend modeling and local morphology preservation; the proposed semantic constraint loss function significantly improves the reconstruction quality of complex areas through the semantically driven local constraint mechanism.

[0008] To achieve the above object, the technical solution created by the present invention is implemented as follows:

[0009] A chain diffusion remote sensing hyperspectral image super-resolution system, comprising:

[0010] The dynamic hypergraph diffusion branch embeds the temporal conditions of the input image into the input image, updates and learns the dynamic adjacency matrix of the obtained features, and obtains affine features based on the dynamic adjacency matrix; denoises and extracts details from the affine features to obtain diffusion features;

[0011] The difference-frequency collaborative attention branch performs a dual-branch spatial attention operation on the input image, completes difference and frequency analysis, and obtains difference features and frequency features respectively; and combines the difference features and frequency features to obtain collaborative features;

[0012] The super-resolution reconstruction branch combines the diffusion features, collaborative features, and the upsampled image of the input image to obtain a super-resolution image;

[0013] Feature fusion is performed between the dynamic hypergraph diffusion branch and the differential-frequency collaborative attention branch through mutual attention operation, and the fused features are then used to guide the dynamic hypergraph diffusion branch and the differential-frequency collaborative attention branch to perform feature processing respectively.

[0014] Furthermore, the dynamic hypergraph diffusion branch includes a plurality of cascaded dynamic hypergraph diffusion unit groups; each dynamic hypergraph diffusion unit group includes two cascaded dynamic hypergraph diffusion units, the output features of the first dynamic hypergraph diffusion unit are input into the mutual attention operation, and the output features of the mutual attention operation are added to the output features of the second dynamic hypergraph diffusion unit by corresponding element-wise addition;

[0015] In each dynamic hypergraph diffusion unit:

[0016] According to the input image or input features, obtain the corresponding basic hypergraph adjacency matrix;

[0017] Add noise to the input image or input feature to obtain the diffusion feature of the initial time step, and determine the corresponding time-varying feature vector based on the number of channels of the input image or input feature; linearly map the time-varying feature vector with the diffusion feature of the current time step to obtain the linear mapping feature; perform query-key attention operation on the linear mapping feature to obtain the similarity matrix;

[0018] Then calculate the Hadamard product of the similarity matrix and the corresponding basic hypergraph adjacency matrix to obtain the corresponding hypergraph adjacency matrix;

[0019] Calculate the product of the corresponding elements of the hypergraph adjacency matrix, linear mapping features, and the learnable projection matrix to obtain the affine features in the current dynamic hypergraph diffusion unit;

[0020] The affine features are denoised and feature extracted, and the processed features are injected into the input image or input features to obtain the diffusion features of the next time step; the diffusion features of the current time step are repeatedly processed on the diffusion features of the next time step to obtain the final diffusion features of the current dynamic hypergraph diffusion unit.

[0021] Furthermore, the process of obtaining the basic hypergraph adjacency matrix in each dynamic hypergraph diffusion unit includes:

[0022] Through the K-nearest neighbor algorithm, the K nodes with the highest spectral similarity to each target node in the input image or input feature are aggregated onto each hyperedge;

[0023] A negative exponential function is used to assign weights to each node on the hyperedge to obtain the basic hypergraph adjacency matrix.

[0024] Furthermore, the process of determining the corresponding time-varying feature vector according to the number of channels of the input image or input feature includes:

[0025] The preliminary time-varying feature vector is generated using sinusoidal time coding by the following formula:

[0026] ;

[0027] ;

[0028] Where k represents the sequence number of the frequency component, represents the fundamental frequency of the kth frequency component, d represents the number of channels of the input image or input feature, represents the preliminary time-varying eigenvector of the t-th iteration;

[0029] Multiple consecutive MLPs are performed on the preliminary time-varying feature vector to perform dimension compression and obtain the time-varying feature vector.

[0030] Furthermore, the process of denoising and extracting features from affine features in each dynamic hypergraph diffusion unit includes: performing a convolution operation on the affine features, performing a group normalization operation on the features after the convolution operation, and performing a GELU activation operation on the features after the group normalization operation to obtain output features.

[0031] Furthermore, the process of injecting the processed features into the input image or input features in each dynamic hypergraph diffusion unit includes:

[0032] ;

[0033] ;

[0034] ;

[0035] in, represents the noise attenuation rate, represents the detail injection rate, Represents the input image or current input features, It represents the features after denoising and feature extraction of affine features, and LinearSchedule represents the function of linearly adjusting hyperparameters.

[0036] Furthermore, in the differential-frequency collaborative attention branch, after convolution feature extraction is performed on the input image to obtain input features, the input features are processed by multiple cascaded differential-frequency collaborative attention unit groups; wherein each differential-frequency collaborative attention unit group includes two cascaded differential-frequency collaborative attention units, the output features of the first differential-frequency collaborative attention unit are input into the mutual attention operation, and the output features of the mutual attention operation are added to the output features of the second differential-frequency collaborative attention unit by corresponding elements; each differential-frequency collaborative attention unit includes a differential attention analysis sub-branch and a frequency analysis sub-branch, and the input features are processed by the differential attention analysis sub-branch and the frequency analysis sub-branch at the same time to obtain differential features and frequency features; the differential features and the frequency features are added by corresponding elements, the added features are subjected to a convolution operation, and then added to the input features by corresponding elements to obtain collaborative features;

[0037] In the differential attention analysis sub-branch, the input features are convolved with multiple sets of spectral difference kernels of different scales to complete the multi-scale spectral dimension sliding calculation difference. The obtained multi-scale spectral features are processed by a multi-layer perceptron to generate multi-scale attention weights. The multi-scale attention weights are fused and then fine-tuned by convolution operation, and then multiplied with the corresponding elements of the input features to obtain differential features.

[0038] In the frequency analysis sub-branch, the Laplace operator and Gaussian kernel are used to extract high-frequency features and low-frequency features from the input features respectively; after the high-frequency features and low-frequency features are channel-spliced, the spliced ​​features are convolutionally fused, and then the fused features are multiplied with the corresponding elements of the input features to obtain the frequency features.

[0039] Furthermore, in the super-resolution reconstruction branch, the corresponding elements of the diffusion features and the collaborative features are added, and then the added features are continuously subjected to convolution and transposed convolution operations, and then the corresponding elements are added to the upsampled image; after the added features are subjected to convolution operations, a super-resolution image is obtained.

[0040] A chain diffusion remote sensing hyperspectral image super-resolution method, comprising:

[0041] S1: Obtain a remote sensing image dataset and preprocess it to obtain a training set;

[0042] S2: inputting the training set obtained in step S1 into the chain diffusion remote sensing hyperspectral image super-resolution system provided by the present invention, and training the super-resolution system in combination with the semantic constraint loss function to obtain a super-resolution model;

[0043] S3: According to the super-resolution model obtained in step S2, adjust the hyperparameters of the training super-resolution system, and repeat step S2 until the optimal super-resolution model is obtained;

[0044] S4: Input the low-resolution remote sensing image into the optimal super-resolution model obtained in step S3 to perform super-resolution reconstruction to obtain a corresponding high-resolution remote sensing image.

[0045] Furthermore, the semantic constraint loss function in step S2 is obtained as follows:

[0046] ;

[0047] in, represents the semantic constraint loss function, represents the L1 loss function, represents the SAM loss function, represents the superpixel nesting loss function, 、 and represents the loss weight;

[0048] Superpixel nesting loss function It is obtained from the following formula:

[0049] ;

[0050] in, represents the spectral loss, represents texture loss, represents the superpixel nesting loss weight;

[0051] Spectral loss It is obtained from the following formula:

[0052] ;

[0053] in, It represents the number of super-pixels obtained by using the SLIC algorithm to perform super-pixel segmentation on the super-resolution image output by the super-resolution model. represents the spectral mean of all pixels in the i-th superpixel in the real high-resolution remote sensing image corresponding to the super-resolution image, Represents the spectral mean of all pixels in the i-th superpixel in the super-resolution image, the spectral mean It is obtained by the following formula:

[0054] ;

[0055] in, represents a superpixel, Indicates the number of pixels contained in the superpixel, Represents the pixel value at position (x, y) in image I;

[0056] Texture loss It is obtained from the following formula:

[0057] ;

[0058] in, represents the texture entropy of all pixels in the i-th superpixel in the real high-resolution remote sensing image, Represents the texture entropy of all pixels in the i-th superpixel in the super-resolution image, texture entropy It is obtained by the following formula:

[0059] ;

[0060] in, represents the probability of the bth bin in the histogram distribution of the LBP value within the superpixel, and B represents the number of bins in the histogram.

[0061] Compared with the prior art, the present invention can achieve the following beneficial effects:

[0062] (1) In the chain-diffusion remote sensing hyperspectral image super-resolution system created by the present invention, dynamic hypergraph learning is combined with the diffusion model, and the iterative optimization process is guided by the time-varying weight space to achieve the joint modeling of the denoising process and feature evolution; through the differential-frequency collaborative attention mechanism, the limitations of the traditional spectral-spatial separation modeling are broken through, and a balance is achieved between global trend modeling and local morphology preservation; the proposed semantic constraint loss function significantly improves the reconstruction quality of complex areas through the semantic-driven local constraint mechanism;

[0063] (2) In the chain-diffusion remote sensing hyperspectral image super-resolution system created by the present invention, the progressive denoising process of the diffusion model is innovatively combined with the dynamic evolution of the hypergraph structure, and the collaborative modeling of noise statistical characteristics and spectral-spatial relationships is achieved through a time-conditioned hyperedge weight adjustment mechanism. The dynamic hypergraph update mechanism allows the model to adaptively adjust the hypergraph node relevance and optimize the learning process according to the characteristic characteristics of different degradation stages. While using noise to supplement information, it simultaneously optimizes the global structural consistency and local detail fidelity, and effectively improves the model's anti-disturbance ability;

[0064] (3) In the chain diffusion remote sensing hyperspectral image super-resolution system created by the present invention, the multi-scale difference strategy can suppress noise interference and retain high-frequency details by fusing the gradient features of different receptive fields (such as local mutations and global trends), thereby accurately maintaining spectral continuity while improving spatial resolution;

[0065] (4) In the chain-diffusion remote sensing hyperspectral image super-resolution method described in the present invention, a superpixel nesting loss is proposed. Superpixels are used as basic semantic units to jointly optimize spectral and texture features, thereby improving spatial resolution while ensuring that the reconstruction results conform to the physical semantic laws of the real world. This design overcomes the shortcomings of traditional pixel-level loss functions in local structure fidelity and semantic consistency, and provides a new optimization perspective for high-dimensional image reconstruction tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The accompanying drawings, which constitute part of the present invention, are intended to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are intended to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings:

[0067] Figure 1 A schematic diagram of the chain-diffusion remote sensing hyperspectral image super-resolution system according to an embodiment of the present invention;

[0068] Figure 2 A schematic diagram of a dynamic hypergraph diffusion unit according to an embodiment of the present invention;

[0069] Figure 3 A schematic diagram of a cascaded differential-frequency collaborative attention unit according to an embodiment of the present invention;

[0070] Figure 4 A schematic diagram of the process of the chain diffusion remote sensing hyperspectral image super-resolution method described in the present invention;

[0071] Figure 5 A visualization comparison of the reconstruction method described in the embodiment of the present invention with other reconstruction methods on the MDAS dataset;

[0072] Figure 6 A visualization comparison of the reconstruction method described in an embodiment of the present invention with other reconstruction methods on the Pavia Centre dataset;

[0073] Figure 7 This is a visualization comparison of the reconstruction method described in the embodiment of the present invention and other reconstruction methods on the Houston dataset. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not constitute a limitation of the present invention.

[0075] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other.

[0076] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In addition, the terms "first", "second" and the like are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, features defined as "first", "second" and the like may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0077] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "installed," "connected," and "connected" should be understood in a broad sense. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to internal connections between two components. Those skilled in the art can understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0078] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments.

[0079] like Figure 1 As shown, the chain-diffusion remote sensing hyperspectral image super-resolution system described in the embodiment of the present invention includes a dynamic hypergraph diffusion branch, a differential-frequency collaborative attention branch, and a super-resolution reconstruction branch. The dynamic hypergraph diffusion branch embeds the time condition of the input image into the input image, updates and learns the obtained features using a dynamic adjacency matrix, and obtains affine features based on the dynamic adjacency matrix; denoises and extracts details on the affine features to obtain diffusion features. The differential-frequency collaborative attention branch performs a dual-branch spatial attention operation on the input image, completes differential and frequency analysis, and obtains differential features and frequency features respectively; and combines the differential features and frequency features to obtain collaborative features. The super-resolution reconstruction branch combines the diffusion features, collaborative features, and an upsampled image of the input image to obtain a super-resolution image. Feature fusion is performed between the dynamic hypergraph diffusion branch and the differential-frequency collaborative attention branch through a mutual attention operation, and the fused features are then used to guide the dynamic hypergraph diffusion branch and the differential-frequency collaborative attention branch to perform feature processing.

[0080] In some embodiments, the dynamic hypergraph diffusion branch includes multiple cascaded dynamic hypergraph diffusion unit groups; each dynamic hypergraph diffusion unit group includes two cascaded dynamic hypergraph diffusion units, the output features of the first dynamic hypergraph diffusion unit are input into the mutual attention operation, and the output features of the mutual attention operation are added to the output features of the second dynamic hypergraph diffusion unit by corresponding elements.

[0081] The structure of each dynamic hypergraph diffusion unit is as follows Figure 2 As shown, according to the input image or input feature, the corresponding basic hypergraph adjacency matrix is ​​obtained. Since there are a large number of mixed pixels in the low-resolution hyperspectral image, the components contained in the mixed pixels with similar spectra usually have a large correlation. Therefore, the present invention constructs a basic hypergraph based on spectral clustering. Specifically, in some embodiments, the process of obtaining the basic hypergraph adjacency matrix in each dynamic hypergraph diffusion unit includes: aggregating the K nodes with the highest spectral similarity with each target node in the input image or input feature onto each hyperedge through the K nearest neighbor algorithm; assigning a weight to each node on the hyperedge through a negative exponential function to obtain the basic hypergraph adjacency matrix.

[0082] In an embodiment of the present invention, the process of assigning a weight to each node on a hyperedge using a negative exponential function can be expressed as follows:

[0083] ;

[0084] in, represents the Euclidean distance between the i-th node and the target node on the hyperedge, represents the average Euclidean distance between all nodes on the hyperedge and the target node; Represents the weight value assigned to the i-th node on the hyperedge;

[0085] The basic hypergraph adjacency matrix is ​​obtained by the following formula:

[0086] ;

[0087] in, represents the basic hypergraph adjacency matrix, Represents the constructed basic hypergraph based on spectral clustering, where the basic hypergraph It is composed of hyperedges, which are composed of the weights of each node constitute, represents the degree of a node in the hypergraph, Represents the degree of the edge and the degree of the node in the hypergraph and edge degree Based on the basic hypergraph Calculated, Represents weight value The matrix is ​​initialized to the identity matrix.

[0088] Add noise to the input image or input feature to obtain the diffusion feature of the initial time step. In this embodiment of the present invention, the input image or input feature Add the noise figure to The standard normal distribution noise is:

[0089] ;

[0090] in, represents the standard normal distribution noise, that is , represents the noise factor, i.e. , Represents the feature after adding noise, feature This is the diffusion characteristic of the initial time step, that is, time step t=0.

[0091] Determining the corresponding time-varying feature vector based on the number of channels of the input image or input feature. In some embodiments, the process of determining the corresponding time-varying feature vector based on the number of channels of the input image or input feature includes:

[0092] The preliminary time-varying feature vector is generated using sinusoidal time coding by the following formula:

[0093] ;

[0094] ;

[0095] Where k represents the sequence number of the frequency component, represents the fundamental frequency of the kth frequency component, d represents the number of channels of the input image or input feature, and the number of channels of the input image is the number of spectral bands. Denotes the preliminary time-varying feature vector at the tth iteration. Multiple consecutive MLPs are performed on the preliminary time-varying feature vector to reduce its dimension and obtain the time-varying feature vector.

[0096] In the embodiment of the present invention, the preliminary time-varying feature vector Perform two consecutive MLP operations, and use the SiLU activation function to activate the features between the two MLP operations to obtain the time-varying feature vector at the tth iteration Among them, the SiLU activation function can significantly improve the feature learning ability of the system through smooth nonlinearity and excellent gradient characteristics. The acquisition process can be expressed as:

[0097] ;

[0098] in, Indicates the first MLP operation The weight of Indicates the second MLP operation The weight of time.

[0099] The time-varying feature vector is combined with the diffusion feature of the current time step to perform linear mapping to obtain the linear mapping feature. Perform broadcast operations to achieve time-varying feature vectors Automatic expansion of the dimension, and then the broadcasted features and the features after adding noise Perform channel splicing to obtain features ,Right now:

[0100] ;

[0101] in, Indicates channel splicing operation, Represents the diffusion characteristics of the current time step;

[0102] The features Perform linear mapping to the unified space and adjust the feature channel to obtain the linear mapping feature ,Right now:

[0103] ;

[0104] in, Represents the weight of the linear mapping.

[0105] The embedding of time vector is to control the inverse diffusion process. Within the iteration of T time steps, each time step t corresponds to a different noise intensity. The embedding vector helps the model distinguish the noise feature distribution at different stages. In the early stage, the noise variance is large, the effective signal is concentrated in the low frequency, and the time-varying feature vector Activate the neurons responsible for low-frequency filtering; in the later stage, high-frequency components are highlighted, and the time-varying eigenvector Activate neurons responsible for high-frequency filtering.

[0106] The query-key attention operation is performed on the linear mapping feature to obtain the similarity matrix. In the embodiment of the present invention, the linear mapping feature of the current time step is observed by the query-key attention mechanism. Similarity matrix ,Right now:

[0107] ;

[0108] ;

[0109] ;

[0110] in, represents the query matrix obtained by the query-key attention operation, represents the query weight, represents the key matrix obtained by the query-key attention operation, represents the key weight, and C represents the linear mapping feature The number of channels.

[0111] Then calculate the Hadamard product of the similarity matrix and the corresponding basic hypergraph adjacency matrix to obtain the corresponding hypergraph adjacency matrix. In an embodiment of the present invention, the hypergraph adjacency matrix is ​​obtained by the following formula:

[0112] ;

[0113] in, represents the hypergraph adjacency matrix, Represents the Hadamard product operation.

[0114] The product of the corresponding elements of the hypergraph adjacency matrix, the linear mapping features, and the learnable projection matrix is ​​calculated to obtain the affine features in the current dynamic hypergraph diffusion unit. In an embodiment of the present invention, the affine features are obtained by the following formula:

[0115] ;

[0116] in, represents the affine feature, Represents the learnable projection matrix, which is initialized by the system and is a system learning parameter.

[0117] Denoising and feature extraction are performed on the affine features, and the processed features are injected into the input image or input features. In some embodiments, the process of denoising and feature extraction on the affine features in each dynamic hypergraph diffusion unit includes: performing a convolution operation on the affine features, performing a group normalization operation on the convolved features, and performing a GELU activation operation on the features after the group normalization operation to obtain output features. In this embodiment of the present invention, the affine features are denoised and extracted. Perform denoising and feature extraction to obtain output features The process can be expressed by the following formula:

[0118] ;

[0119] Among them, Conv represents the two-dimensional convolution operation, and GN represents the group normalization operation (GroupNorm). Group normalization can effectively eliminate the dependence on batch size and enhance task generalization.

[0120] In some embodiments, the process of injecting the processed features into the input image or input features in each dynamic hypergraph diffusion unit to obtain the diffusion features of the next time step includes:

[0121] ;

[0122] ;

[0123] ;

[0124] in, represents the noise attenuation rate, Represents the detail injection rate, and LinearSchedule represents the function of linearly adjusting hyperparameters. This function is a tool for linearly adjusting hyperparameters in machine learning.

[0125] Repeat the process of the diffusion feature of the current time step for the diffusion feature of the next time step to obtain the final diffusion feature of the current dynamic hypergraph diffusion unit. Repeat the above diffusion characteristics for the current time step The processing operation is performed, and T time steps are iterated to obtain the final diffusion characteristics of the current dynamic hypergraph diffusion unit .

[0126] In some embodiments, in the differential-frequency collaborative attention branch, after convolution feature extraction is performed on the input image to obtain input features, the input features are processed by multiple cascaded differential-frequency collaborative attention unit groups; wherein, each differential-frequency collaborative attention unit group includes two cascaded differential-frequency collaborative attention units, and the output features of the first differential-frequency collaborative attention unit are input into the mutual attention operation, and the output features of the mutual attention operation are added to the output features of the second differential-frequency collaborative attention unit by corresponding elements.

[0127] In the embodiment of the present invention, the input feature is obtained by the following formula:

[0128] ;

[0129] in, Represents input features, I represents input image, and ReLU represents ReLU activation operation.

[0130] The differential-frequency collaborative attention branch includes two cascaded differential-frequency collaborative attention unit groups, each differential-frequency collaborative attention unit group includes two cascaded differential-frequency collaborative attention units, and the output features of the first differential-frequency collaborative attention unit are input into the mutual attention operation, combined with the output features of the first dynamic hypergraph diffusion unit to input into the mutual attention operation. It can be understood that the input of the mutual attention operation includes the collaborative features output by the differential-frequency collaborative attention branch and the diffusion features output by the dynamic hypergraph diffusion branch. In an embodiment of the present invention, the mutual attention operation is represented by the following formula:

[0131] ;

[0132] in, represents the output features of the mutual attention operation, The query matrix representing the diffusion features, The bond matrix representing the collaborative features, represents the value matrix of the collaborative features, represents the mutual attention operation.

[0133] The structure of each difference-frequency collaborative attention unit is as follows Figure 3 As shown in the figure, it includes a differential attention analysis sub-branch and a frequency analysis sub-branch. The input features are processed by the differential attention analysis sub-branch and the frequency analysis sub-branch at the same time to obtain differential features and frequency features respectively. The differential features and frequency features are added together by corresponding elements. After the added features are convolved, they are added together with the input features by corresponding elements to obtain collaborative features.

[0134] In an embodiment of the present invention, residual processing is performed on the input features, specifically, the basic residual block in the classic ResNet network is used to perform residual processing on the input features, and the features after residual processing are added to the corresponding elements of the features after addition and convolution to obtain collaborative features.

[0135] The processing of each difference-frequency co-attention unit can be expressed as follows:

[0136] ;

[0137] in, represents collaborative features, Represents the features after residual processing of the input features, represents the frequency characteristics, Represents differential features.

[0138] In the differential attention analysis sub-branch, the input features are convolved with multiple sets of spectral difference kernels of different scales to perform sliding differential calculations on the multi-scale spectral dimensions. The resulting multi-scale spectral features are processed using a multi-layer perceptron to generate multi-scale attention weights. These multi-scale attention weights are fused and fine-tuned using a convolution operation before being multiplied by the corresponding elements of the input features to generate differential features.

[0139] In the embodiment of the present invention, the input feature One-dimensional convolution is performed using spectral difference kernels of [-1, 1, 0], [-1, 0, 1], and [-1, -1, 0, 1, 1], namely:

[0140] ;

[0141] ;

[0142] ;

[0143] in, 、 and Respectively represent the spectral features after convolution of three groups of spectral difference kernels, Indicates sliding calculation of differences along the spectral dimension;

[0144] The spectral characteristics , spectral characteristics and spectral characteristics Perform channel average pooling separately to obtain the global statistics of differential information, namely:

[0145] ;

[0146] in, represents the jth spectral feature after pooling, j = 1, 2, 3, H and W represent the height and width of the feature respectively;

[0147] Input features Channel average pooling is also performed to obtain input features The global statistics of , namely:

[0148] ;

[0149] in, Represents the spectral features after pooling.

[0150] The obtained multi-scale spectral features 、 、 and Generate multi-scale attention weights through the bottleneck structure of the multi-layer perceptron MLP 、 、 and ,Right now:

[0151] ;

[0152] After the multi-scale attention weights are fused by adding the corresponding elements, a 1×1 convolution fine-tuning operation is performed, that is,

[0153] ;

[0154] in, represents the weight after fusion, Represents the Sigmoid activation operation, Represents a 1×1 two-dimensional convolution operation;

[0155] The weight With input features Multiply the corresponding elements to get the differential features ,Right now:

[0156] .

[0157] In the frequency analysis sub-branch, the Laplace operator and Gaussian kernel are used to extract high-frequency features and low-frequency features from the input features respectively; after the high-frequency features and low-frequency features are channel-spliced, the spliced ​​features are convolutionally fused, and then the fused features are multiplied with the corresponding elements of the input features to obtain the frequency features.

[0158] In the embodiment of the present invention, the Laplace operator and the Gaussian kernel are used to extract the input features from the following formula: Extract high-frequency features and low-frequency characteristics ,Right now:

[0159] ;

[0160] ;

[0161] Among them, Laplacian represents the Laplace operator, Gaussian represents the Gaussian kernel, Represents a 3×3 two-dimensional convolution operation;

[0162] The high-frequency features and low-frequency characteristics After channel splicing, convolution fusion operation is performed on the spliced ​​features, and then the fused features are combined with the input features. Multiply the corresponding elements to get the frequency characteristics ,Right now:

[0163] .

[0164] In some embodiments, as Figure 1 As shown in the figure, in the super-resolution reconstruction branch, the corresponding elements of the diffusion feature and the collaborative feature are added, and then the added features are continuously convolved and transposed convolution operations are performed, and then the corresponding elements are added to the upsampled image; after the added features are convolved, a super-resolution image is obtained.

[0165] In the super-resolution reconstruction branch of the embodiment of the present invention, the diffusion feature and collaborative features The corresponding elements are added, and then the added features are continuously convolved and transposed convolution operations are performed, that is:

[0166] ;

[0167] in, represents the transposed convolution operation, Indicates the currently obtained features;

[0168] The features The corresponding elements are added to the bicubic upsampled image of the input image, and the added features are convolved to obtain the super-resolution image. ,Right now:

[0169] ;

[0170] in, An image representing the bicubic upsampled version of the input image.

[0171] The present invention also provides a chain diffusion remote sensing hyperspectral image super-resolution method, such as Figure 4 As shown, the method includes:

[0172] S1: Obtain a remote sensing image dataset and preprocess the remote sensing image dataset to obtain a training set.

[0173] In one embodiment of the present invention, the process of preprocessing a remote sensing image dataset includes: segmenting the remote sensing hyperspectral image dataset, selecting 70% of the image area as the training portion of the training set, and 30% of the area as the validation portion of the training set. Randomly selecting patches from each area of ​​the remote sensing image, setting different patch sizes based on the magnification factor, where magnifications of 2, 3, and 4 times correspond to patch sizes of 64×64, 96×96, and 128×128, respectively. Each patch is randomly horizontally flipped, rotated at different angles, and scaled by different magnifications. Finally, bicubic downsampling is used to downsample these patches to low-resolution 32×32 hyperspectral images according to different scaling factors to obtain the training set.

[0174] S2: Input the training set obtained in step S1 into the chain diffusion remote sensing hyperspectral image super-resolution system provided by the present invention, and train the super-resolution system in combination with the semantic constraint loss function to obtain a super-resolution model.

[0175] During the training process, the training part of the training set completes the training of the super-resolution system, and the verification part of the training set completes the calculation and verification of the semantic constraint loss function of the current trained model, and adjusts the weight parameters in the super-resolution model based on the feedback of the semantic constraint loss function to the super-resolution model.

[0176] In some embodiments, the semantic constraint loss function is given by:

[0177] ;

[0178] in, represents the semantic constraint loss function, represents the L1 loss function, represents the SAM loss function, represents the superpixel nesting loss function, 、 and Represents the loss weight, which is adaptively adjusted according to the actual training situation. In this embodiment of the present invention, the weight 、 and 1, 0.2, and 0.2 respectively

[0179] Superpixel nesting loss function It is obtained from the following formula:

[0180] ;

[0181] in, represents the spectral loss, represents texture loss, Represents the superpixel nesting loss weight, which is adaptively adjusted according to the actual training situation;

[0182] Spectral loss It is obtained from the following formula:

[0183] ;

[0184] in, It represents the number of super-pixels obtained by using the SLIC algorithm to perform super-pixel segmentation on the super-resolution image output by the super-resolution model. represents the spectral mean of all pixels in the i-th superpixel in the real high-resolution remote sensing image corresponding to the super-resolution image, Represents the spectral mean of all pixels in the i-th superpixel in the super-resolution image, the spectral mean It is obtained by the following formula:

[0185] ;

[0186] in, represents a superpixel, Indicates the number of pixels contained in the superpixel, Represents the pixel value at position (x, y) in image I;

[0187] Texture loss It is obtained from the following formula:

[0188] ;

[0189] in, represents the texture entropy of all pixels in the i-th superpixel in the real high-resolution remote sensing image, Represents the texture entropy of all pixels in the i-th superpixel in the super-resolution image, texture entropy It is obtained by the following formula:

[0190] ;

[0191] in, represents the probability of the bth bin in the histogram distribution of the LBP value within the superpixel, and B represents the number of bins in the histogram.

[0192] S3: According to the super-resolution model obtained in step S2, the hyperparameters of the training super-resolution system are adjusted, and step S2 is repeated until the optimal super-resolution model is obtained.

[0193] In the embodiment of the present invention, the optimizer selects Adam, in which the first-order moment exponential attenuation factor , the second-order moment exponential decay factor , with a batch size of 4 and an initial learning rate of 0.0001. The number of training iterations was set to 40. The model parameters were updated based on the semantic constraint loss function. As the semantic constraint loss function gradually decreased, the model's super-resolution reconstruction of low-resolution images improved. After completing several rounds of network parameter updates, the model's performance was tested using the validation set from the training dataset. Training ended when the objective image quality evaluation indicators of the reconstructed images in the validation set stopped improving or reached the preset upper limit of training iterations.

[0194] S4: Input the low-resolution remote sensing image into the optimal super-resolution model obtained in step S3 to perform super-resolution reconstruction to obtain a corresponding high-resolution remote sensing image.

[0195] In order to clearly demonstrate the good super-resolution reconstruction effect of the chain diffusion remote sensing hyperspectral image super-resolution system and method provided by the present invention, the embodiment of the present invention adopts peak signal-to-noise ratio (PSNR), structural similarity (SSIM), spectral angle similarity (SAM) and relative global error separation metric (ERGAS) to objectively evaluate the quality of reconstructed remote sensing hyperspectral images, and uses bicubic interpolation (Bicubic), 3D-FCNN (published in the paper "Hyperspectral Image Spatial Super-Resolution via 3D Full Convolutional Neural Network" in the journal "Remote Sensing"), GDRRN (published in the paper "Single Hyperspectral Image Super-resolution with Grouped Deep Recursive Residual Network" in the journal "2018 IEEE Fourth International Conference on Multimedia Big Data (BigMM)"), SSPSR (published in the paper "Learning Spatial-Spectral Prior for Super-Resolution of Hyperspectral" in the journal "IEEE Transactions on Computational Imaging"), and the like. Imagery”), MCNet (published in the journal “Remote Sensing”, “Mixed 2D / 3D Convolutional Network for Hyperspectral Image Super-Resolution”), ERCSR (published in the journal “IEEE Transactions on Geoscience and Remote Sensing”, “Exploring the Relationship Between 2D / 3D Convolution for Hyperspectral Image Super-Resolution”), GELIN (published in the journal “IEEE Transactions on Geoscience and Remote Sensing”, “A Group-Based Embedding Learning and Integration Network for Hyperspectral ImageSuper-Resolution”), SRDNet (published in the journal “IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING”, “Hyperspectral Image Super-Resolution via Dual-Domain Network Based on Hybrid Convolution”), MSSR (published in the journal “IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING”, “Remote Sensing Hyperspectral Image Super-Resolution via Multidomain SpatialInformation and Multiscale Spectral Information Fusion”), CST method (published in the journal “IEEE TRANSACTIONS IMAGE PROCESS”, “Cross-Scope Spatial-SpectralInformation Aggregation for Hyperspectral Image Super-Resolution”) and the method provided by the present invention (CDiff-HG) were respectively used in the MDAS dataset, PaviaCentre dataset and Houston dataset. The objective comparison results of the method provided by the present invention and the comparative method on the MDAS dataset are shown in Table 1. The present invention achieved the best overall performance at all scaling factors, validating its effectiveness in joint modeling of spatial-spectral features. From an objective perspective, the present invention performed outstandingly in two key metrics: PSNR and SSIM. Taking ×2 super-resolution reconstruction as an example, the PSNR of the present invention reached 34.487dB, a 0.165dB improvement over the suboptimal model MSSR (34.322dB); the SSIM was 0.9244, superior to MSSR's 0.9215. This advantage remained significant at higher scaling factors (×3 and ×4), demonstrating the model's strong ability to recover high-frequency details. Furthermore, the present invention also performed well in the spectral fidelity metric SAM and the error sensitivity metric ERGAS, validating its ability to suppress spectral distortion. Compared to existing methods, our proposed method achieves a PSNR of 31.576 dB in the ×3 super-resolution task, improving by 0.146 dB and 0.125 dB over SRDNet (31.430 dB) and MSSR (31.451 dB), respectively. Its SSIM (0.8524) significantly outperforms other models. Notably, in ×4 super-resolution, despite a smaller PSNR improvement, SAM and ERGAS maintain the highest performance. This demonstrates that our proposed method maintains spectral continuity while more accurately controlling reconstruction error. This property is particularly important for subsequent hyperspectral image interpretation. While existing methods improve performance through 3D convolution or Transformer architectures, their ability to model global trends and local morphology in spectral dimensions remains limited. Our proposed method, by introducing dynamic hypergraph learning and a diffusion model, explicitly models the local correlations and global distribution characteristics of spatial-spectral features in hyperspectral imagery, resulting in enhanced robustness in complex degradation scenarios. Figure 5 Visualization results from a ×4 super-resolution task are shown. Overall, the reconstructed image by this method exhibits clearer outlines and more detailed information in areas with complex textures. For example, in the annotated areas of the visualized image, the restored houses and roads exhibit more realistic structures and less blur.

[0196] Table 1:

[0197]

[0198] Table 2 shows objective comparison results of the proposed method and comparative methods on the Pavia Centre dataset. The proposed method demonstrates significant advantages across all scaling factors, particularly in balancing spatial detail recovery and spectral fidelity. In the ×2 super-resolution task, the proposed method achieves the best PSNR (36.376 dB) and SSIM (0.9565), respectively improving by 0.140 dB and 0.0013 over MSSR (36.236 dB and 0.9552). Its SAM and ERGAS values ​​are also the lowest, demonstrating that the model effectively suppresses spectral distortion while improving spatial resolution. Notably, when the scaling factor is increased to ×3, the proposed method achieves even more significant improvements in PSNR (31.807 dB) and SSIM (0.8896) compared to MSSR (31.547 dB and 0.8836), reaching 0.260 dB and 0.0060, respectively, further validating the robustness of the proposed method. In the challenging ×4 super-resolution task, the proposed method surpasses all compared models in both PSNR and SSIM, with PSNR improving by 0.117dB and SSIM improving by 0.0041 over the next-best model, SSPSR. This method constructs a joint spatial-spectral representation through a hypergraph diffusion unit and a differential-frequency collaborative attention mechanism. Combined with a superpixel nested loss function, it can adaptively capture local structural features and global distribution patterns at different scales. The proposed method's superior performance on the Pavia Centre dataset further validates its cross-dataset generalization capabilities. Figure 6 Visualization results for the ×4 super-resolution task are shown. Compared to earlier models, the images reconstructed by MSSR, CST, and our method have richer textures. In the annotated areas, only our method can fully restore the original building edge shapes.

[0199] Table 2:

[0200]

[0201] Table 3 shows objective comparison results of the proposed method and comparative methods on the Houston dataset. The proposed method achieves optimal results in the ×2, ×3, and ×4 super-resolution tasks, further validating its generalization and stability across different scenarios. In the ×2 super-resolution task, the proposed method significantly outperforms other models in terms of PSNR (38.226dB) and SSIM (0.9506), respectively improving by 0.139dB and 0.0016 over MSSR (38.087dB and 0.9490). Its SAM (2.520) and ERGAS (5.955) also reach the lowest values, demonstrating that the model effectively enhances spatial detail while preserving spectral fidelity. Compared to the Transformer-based CST model, the proposed method achieves a 0.323dB improvement in PSNR and a 4.26% reduction in SAM, demonstrating the superiority of dynamic hypergraph learning in modeling spectral correlations. Similarly, the proposed method maintains a stable lead in the ×3 and ×4 super-resolution tasks. This shows that its progressive optimization strategy can more effectively suppress the accumulation of reconstruction errors, highlighting the strong robustness of the model for spectral-spatial reconstruction under different scaling factors. Figure 7 Visualization results for the ×4 super-resolution task are shown. Overall, CST and our method appear to have the best performance, particularly in effectively mitigating blurring. In the annotated areas of the visualization results, only our method recovers coherent and angularly accurate edges compared to the other methods.

[0202] Table 3:

[0203]

[0204] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in the present disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved. This is not limited herein.

[0205] The above specific embodiments do not limit the scope of protection of the present invention. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A chain diffusion remote sensing hyperspectral image super-resolution system, characterized by: include: A dynamic hypergraph diffusion branch embeds the temporal conditions of the input image into the input image, updates and learns the dynamic adjacency matrix of the obtained features, and obtains affine features based on the dynamic adjacency matrix; denoises and extracts details from the affine features to obtain diffusion features; The difference-frequency collaborative attention branch performs a dual-branch spatial attention operation on the input image to complete difference and frequency analysis to obtain difference features and frequency features respectively; Combining the differential feature and the frequency feature to obtain a collaborative feature; a super-resolution reconstruction branch, combining the diffusion feature, the collaborative feature, and an upsampled image of the input image to obtain a super-resolution image; Feature fusion is performed between the dynamic hypergraph diffusion branch and the differential-frequency collaborative attention branch through mutual attention operation, and the fused features are then used to guide the dynamic hypergraph diffusion branch and the differential-frequency collaborative attention branch to perform feature processing respectively.

2. The chain diffusion remote sensing hyperspectral image super-resolution system according to claim 1, characterized in that: The dynamic hypergraph diffusion branch includes a plurality of cascaded dynamic hypergraph diffusion unit groups; each dynamic hypergraph diffusion unit group includes two cascaded dynamic hypergraph diffusion units, the output features of the first dynamic hypergraph diffusion unit are input into the mutual attention operation, and the output features of the mutual attention operation are added to the output features of the second dynamic hypergraph diffusion unit by corresponding element-wise addition; In each dynamic hypergraph diffusion unit: According to the input image or input features, obtain the corresponding basic hypergraph adjacency matrix; Add noise to the input image or input feature to obtain the diffusion feature of the initial time step, and determine the corresponding time-varying feature vector according to the number of channels of the input image or input feature; linearly map the time-varying feature vector to the diffusion feature of the current time step to obtain a linear mapping feature; Performing a query-key attention operation on the linear mapping feature to obtain a similarity matrix; Then, the Hadamard product of the similarity matrix and the corresponding basic hypergraph adjacency matrix is ​​calculated to obtain the corresponding hypergraph adjacency matrix; Calculating the product of corresponding elements of the hypergraph adjacency matrix, the linear mapping features, and the learnable projection matrix to obtain the affine features in the current dynamic hypergraph diffusion unit; Denoising and feature extraction are performed on the affine features, and the processed features are injected into the input image or input features to obtain the diffusion features of the next time step; The diffusion features of the next time step are processed repeatedly as the diffusion features of the current time step, and multiple time steps are iterated to obtain the final diffusion features of the current dynamic hypergraph diffusion unit.

3. The chain diffusion remote sensing hyperspectral image super-resolution system according to claim 2, characterized in that: The process of obtaining the basic hypergraph adjacency matrix in each dynamic hypergraph diffusion unit includes: Through the K-nearest neighbor algorithm, the K nodes with the highest spectral similarity to each target node in the input image or input feature are aggregated onto each hyperedge; A weight is assigned to each node on the hyperedge through a negative exponential function to obtain the basic hypergraph adjacency matrix.

4. The chain diffusion remote sensing hyperspectral image super-resolution system according to claim 2, characterized in that: The process of determining the corresponding time-varying feature vector according to the number of channels of the input image or input feature includes: The preliminary time-varying feature vector is generated using sinusoidal time coding by the following formula: ； ； Among them, k represents the serial number of the frequency component, represents the fundamental frequency of the kth frequency component, d represents the number of channels of the input image or input feature, represents the preliminary time-varying eigenvector of the t-th iteration; A plurality of continuous MLPs are performed on the preliminary time-varying feature vector to perform dimension compression to obtain the time-varying feature vector.

5. The chain diffusion remote sensing hyperspectral image super-resolution system according to claim 2, characterized in that: The process of denoising and extracting features from the affine features in each dynamic hypergraph diffusion unit includes: performing a convolution operation on the affine features, performing a group normalization operation on the features after the convolution operation, and performing a GELU activation operation on the features after the group normalization operation to obtain output features.

6. The chain diffusion remote sensing hyperspectral image super-resolution system according to claim 2, characterized in that: The process of injecting the processed features into the input image or input features in each dynamic hypergraph diffusion unit includes: ； ； ； in, represents the noise attenuation rate, represents the detail injection rate, Represents the input image or current input features, represents the features after denoising and feature extraction are performed on the affine features, and LinearSchedule represents a function for linearly adjusting hyperparameters.

7. The chain diffusion remote sensing hyperspectral image super-resolution system according to claim 1, characterized in that: In the differential-frequency collaborative attention branch, after convolution feature extraction is performed on the input image to obtain input features, the input features are processed by multiple cascaded differential-frequency collaborative attention unit groups; wherein, each differential-frequency collaborative attention unit group includes two cascaded differential-frequency collaborative attention units, and the output features of the first differential-frequency collaborative attention unit are input into the mutual attention operation, and the output features of the mutual attention operation are added to the output features of the second differential-frequency collaborative attention unit by corresponding elements; each differential-frequency collaborative attention unit includes a differential attention analysis sub-branch and a frequency analysis sub-branch, and the input features are processed by the differential attention analysis sub-branch and the frequency analysis sub-branch at the same time to obtain the differential features and the frequency features; the differential features and the frequency features are added by corresponding elements, and the added features are subjected to a convolution operation and then added to the input features by corresponding elements to obtain the collaborative features; In the differential attention analysis sub-branch, the input features are convolved using multiple sets of spectral difference kernels of different scales to complete the multi-scale spectral dimension sliding calculation difference; the obtained multi-scale spectral features are processed by a multi-layer perceptron to generate multi-scale attention weights; the multi-scale attention weights are fused and then fine-tuned by convolution, and then multiplied with the corresponding elements of the input features to obtain the differential features; In the frequency analysis sub-branch, the Laplace operator and the Gaussian kernel are used to extract high-frequency features and low-frequency features from the input features respectively; after the high-frequency features and the low-frequency features are channel-spliced, a convolution fusion operation is performed on the spliced ​​features, and then the fused features are multiplied by the corresponding elements of the input features to obtain the frequency features.

8. The chain diffusion remote sensing hyperspectral image super-resolution system according to claim 1, characterized in that: In the super-resolution reconstruction branch, the corresponding elements of the diffusion feature and the collaborative feature are added, and then the added features are continuously subjected to convolution and transpose convolution operations, and then the corresponding elements are added to the upsampled image; after the added features are subjected to convolution operations, the super-resolution image is obtained.

9. A chain diffusion remote sensing hyperspectral image super-resolution method, characterized in that: include: S1: Acquire a remote sensing image dataset and preprocess the remote sensing image dataset to obtain a training set; S2: inputting the training set obtained in step S1 into the chain diffusion remote sensing hyperspectral image super-resolution system according to any one of claims 1 to 8, and training the super-resolution system in combination with a semantic constraint loss function to obtain a super-resolution model; S3: According to the super-resolution model obtained in step S2, the hyperparameters of the super-resolution system are adjusted and trained, and step S2 is repeated until an optimal super-resolution model is obtained; S4: Input the low-resolution remote sensing image into the optimal super-resolution model obtained in step S3 to perform super-resolution reconstruction to obtain a corresponding high-resolution remote sensing image.

10. The chain diffusion remote sensing hyperspectral image super-resolution method according to claim 9, characterized in that: The semantic constraint loss function in step S2 is obtained as follows: ； in, represents the semantic constraint loss function, represents the L1 loss function, represents the SAM loss function, represents the superpixel nesting loss function, 、 and represents the loss weight; Superpixel nesting loss function It is obtained from the following formula: ； in, represents the spectral loss, represents texture loss, represents the superpixel nesting loss weight; Spectral loss It is obtained from the following formula: ； in, represents the number of superpixels obtained by performing superpixel segmentation on the super-resolution image output by the super-resolution model using the SLIC algorithm, represents the spectral mean of all pixels in the i-th superpixel in the real high-resolution remote sensing image corresponding to the super-resolution image, represents the spectral mean of all pixels in the i-th superpixel in the super-resolution image. It is obtained by the following formula: ； in, represents a superpixel, represents the number of pixels contained in the superpixel, Represents the pixel value at position (x, y) in image I; Texture loss It is obtained from the following formula: ； in, represents the texture entropy of all pixels in the i-th superpixel in the real high-resolution remote sensing image, Represents the texture entropy of all pixels in the i-th superpixel in the super-resolution image, texture entropy It is obtained by the following formula: ； in, represents the probability of the bth bin in the histogram distribution of the LBP value within the superpixel, and B represents the number of bins in the histogram.

Images