Cross-modal generative star-ground network semantic communication method and system with super compression rate

Abstract

The application relates to a cross-modal generative satellite-ground network semantic communication method with super-high compression rate, which comprises the following steps: S1, establishing a generative cross-modal satellite-ground network semantic communication system; S2, establishing a satellite-ground network semantic communication system model fusing NOMA technology; S3, establishing a dynamic switching algorithm; S4, establishing a double-state satellite-ground link simulation model: considering the influence of the multipath, shadow and Doppler factors existing in satellite-ground communication, dividing the satellite-ground transmission link into two states of good and bad, modeling the channel characteristics by using Rician and Rayleigh distribution, and simulating the Doppler frequency shift through Doppler phase rotation, and completing satellite-ground channel modeling and simulation. In the single-user scene, the original image information is efficiently compressed into a compact text semantic representation through the cross-modal semantic compression module, an end-to-end generative model is adopted to guarantee high semantic consistency and robust reconstruction, the data volume is greatly compressed, and low-delay semantic transmission is achieved, so that the bandwidth burden of the satellite-ground network is effectively reduced.

Description

Technical Field

[0001] This invention belongs to the field of information and communication technology, specifically relating to a cross-modal generative satellite-to-ground network semantic communication method and system with ultra-high compression ratio. Background Technology

[0002] With the continuous advancement of sixth-generation mobile communication technology, integrated space-ground networks, capable of achieving seamless global coverage, are gradually becoming an important component of future communication systems. Their wide-area coverage and high mobility make them indispensable in critical application scenarios such as emergency communication, remote sensing, and intelligent transportation. However, space-ground communication faces severe challenges, including harsh and time-varying channel environments, limited bandwidth resources, and complex and diverse mission requirements, placing immense pressure on traditional transmission mechanisms.

[0003] In recent years, semantic communication has emerged as a new paradigm that breaks through the bottleneck of traditional bit-level precision transmission. Its core idea is to extract, transmit, and reconstruct high-level semantic information related to the task, placing semantic correctness above traditional bit-level accuracy, thereby achieving more efficient and robust communication under extreme channel conditions.

[0004] However, truly applying semantic communication technology to space-ground converged networks faces three major challenges: First, due to the scarcity of satellite spectrum resources and the limited and non-renewable bandwidth, traditional methods for transmitting large amounts of image data cannot meet the demands for efficient propagation. Furthermore, most existing research only considers single-user access scenarios, making it difficult to address the challenges of dense access by a large number of users. Second, space-ground links are susceptible to interference from complex environments such as shadowing, multipath fading, and the Doppler effect during actual transmission, leading to decreased signal robustness and reliability. Third, space-ground converged network services involve diverse and conflicting tasks; for example, low latency and high fidelity are mutually restrictive factors, and existing semantic communication methods can mostly only optimize for single-task requirements, failing to simultaneously meet multiple tasks. Therefore, there is an urgent need for a communication method that can simultaneously improve the transmission quality of space-ground channels, reduce the data transmission volume of space-ground channels, solve the problem of simultaneous access by a large number of users, and meet the needs of multiple tasks. Summary of the Invention

[0005] To address the challenges of harsh and time-varying environments, limited bandwidth resources, and complex and diverse mission requirements in existing satellite-to-ground network communication technologies, this invention proposes a cross-modal generative satellite-to-ground network semantic communication method and system with ultra-high compression ratio.

[0006] The present invention provides a cross-modal generative satellite-to-ground network semantic communication method with ultra-high compression ratio, comprising the following steps:

[0007] S1. Establish a generative cross-modal satellite-to-ground network semantic communication system: For single-user scenarios, the image data to be transmitted is transformed into a compact text semantic representation in the cross-modal semantic compression module, and then transmitted through the satellite-to-ground channel module. At the receiving end, the cross-modal generative reconstruction module uses a text-driven end-to-end generative model to reconstruct the image, achieving high semantic consistency, efficient compression, and robust communication.

[0008] S2. Establish a semantic communication system model for satellite-to-ground network integrating NOMA technology: For multi-user scenarios, under the NOMA mechanism, the signals of multiple users are superimposed in the power domain, and serial interference cancellation technology is used at the receiving end to separate the signals of each user.

[0009] S3. Establish a dynamic switching algorithm: Based on the real-time signal-to-noise ratio of the channel and the requirements of the task type, adaptively select the text compression-generative reconstruction path or the high-fidelity image direct transmission path to achieve a dynamic balance between bandwidth efficiency, semantic consistency and image fidelity.

[0010] S4. Establish a dual-state satellite-to-ground link simulation model: Taking into account the effects of multipath, shadowing and Doppler factors in satellite-to-ground communication, the satellite-to-ground transmission link is divided into two states: good and bad. The channel characteristics are modeled using Rician and Rayleigh distributions, and Doppler frequency shift is simulated by Doppler phase rotation to complete the satellite-to-ground channel modeling and simulation.

[0011] Furthermore, in S1, the specific method of the generative cross-modal satellite-to-ground network semantic communication system includes:

[0012] S11, Cross-modal semantic compression module

[0013] In the feature extraction stage, the BLIP-2 architecture is adopted, combining a frozen visual encoder, a frozen large language model, and a trainable intermediate module—a query transformer—to construct a visual-language bridging path; the transmitted image is... , and These represent the height and width of the image, respectively. After passing through the ViT image encoder, a set of image feature vectors will be generated:

[0014]

[0015] Where N is the number of blocks in the image, and d is the feature dimension of the block;

[0016] Introduce M learnable query vectors As input to the Q-Former, it interacts with itself using a self-attention mechanism and is fused with image features V through cross-attention to extract the semantic information most crucial for text generation:

[0017]

[0018] BLIP-2 includes a two-stage pre-training strategy. The first stage optimizes the Q-Former by jointly performing image-text comparison, image caption generation, and matching tasks, learning how to extract strongly correlated representations from visual features. The second stage passes the output of the Q-Former through a projection layer. Input space required for projection to LLM:

[0019]

[0020] The visual guidance cues are prefixed to the text sequence and input into the frozen LLM, guiding it to generate a natural language description corresponding to the image. ,in Indicates the i-th word, For the vocabulary list:

[0021]

[0022] After text generation, BERT is used to perform semantic embedding encoding on the natural language sequence. This model, based on the Transformer architecture, captures semantic information in the text through bidirectional context, encoding the text sequence T into a context-sensitive vector representation. ,Right now ;

[0023] S12, Satellite-to-Ground Channel Transmission Module

[0024] A unified simulation model was performed for the satellite-to-ground channel, taking into account path loss, Doppler shift, and shadowing fading factors.

[0025]

[0026] in, For satellite-to-ground channel gain, It is additive Gaussian noise.

[0027] Furthermore, in S1, the specific method of the generative cross-modal satellite-to-ground network semantic communication system also includes:

[0028] S13, Cross-modal generative reconstruction module

[0029] First, the system maps the received channel output Y into a natural language description using a text inverse encoder, Transformer. This process is autoregressive:

[0030]

[0031] The generated text The core semantic components of the image are preserved, and can be used for subsequent image generation;

[0032] In the image reconstruction stage, a text-driven diffusion-based image generation model, Stable Diffusion, is employed. This model learns the image distribution under textual conditions in the latent space and achieves high-quality image generation through an anti-diffusion process: First, the text is... The input text encoder obtains the condition vector. Then Gaussian noise T-step anti-diffusion is performed:

[0033]

[0034] During the generation process, a consistent random seed and a fixed inference step size are used to control the diversity and stability of the images;

[0035] Meanwhile, GCM-SeC introduces multiple loss functions during the training phase, including semantic consistency scores. Mean square error loss Cross-entropy loss The specific optimization objective is to minimize the following reconstruction loss:

[0036]

[0037] in , , The corresponding weights.

[0038] Furthermore, in S2, the satellite-to-ground network semantic communication system model incorporating NOMA technology specifically includes:

[0039] S21, Cross-modal semantic compression module

[0040] For each user, the channel is obtained using the cross-modal semantic compression module in GCM-SeC. For image S,

[0041] First, the image semantic feature extraction module transforms the text semantic representation T into a corresponding text semantic representation. Then, the joint source-channel encoder module transforms the text representation T into a vector that the channel needs to transmit. ,Right now

[0042]

[0043] in, θ represents the joint source-channel coding, θ represents the parameters of the neural network, and SNR is the signal-to-noise ratio;

[0044] S22, Power Superposition

[0045] For each user The semantic encoding represents power weighting, which is then superimposed on the power domain to form a composite signal:

[0046]

[0047] in, The power factor allocated to the i-th user satisfies .

[0048] Furthermore, in S2, the satellite-to-ground network semantic communication system model incorporating NOMA technology specifically includes:

[0049] S23, Channel Transmission

[0050] The resulting composite signal The signal will be transmitted via satellite-to-ground link. :

[0051]

[0052] in, For satellite-to-ground channel gain, It is additive Gaussian noise;

[0053] S24, Serial interference cancellation

[0054] On the receiver side, the system uses continuous interference to cancel out the interference on the received signal. Decoding is performed; in a dual-user scenario, the receiver first decodes the signal from the user with higher power. Decoding is performed, treating signals from other users as interference; after decoding, the signal is reconstructed and subtracted from the composite signal, thereby recovering the signal from the lower-power user. :

[0055]

[0056] S25, Cross-modal generative reconstruction module

[0057] For the decoded signals of each user First, the signal received from the channel is obtained through a joint source-channel decoder. Transform into the corresponding text semantic representation ,Right now

[0058]

[0059] in, Indicates joint source-channel decoding, The parameters of the neural network are then used by the image generative reconstruction module to convert them into a reconstructed image. .

[0060] Furthermore, in S3, the dynamic switching algorithm specifically includes:

[0061] S31. Channel and mission status parameter input

[0062] Real-time channel signal-to-noise ratio of received input and task requirement vector Each dimension of the task vector These correspond to specific demand indicators in satellite-to-ground scenarios.

[0063] S32. Branch scoring function construction: Based on the current channel state and task vector, set the SNR range. and A comprehensive scoring function is established by normalizing the signal-to-noise ratio and weighting the task score:

[0064]

[0065] parameter , Used to adjust the relative weights of channel evaluation and mission requirements in branch decisions. The importance weights of each dimension of requirements can be flexibly configured according to the task scenario, within the task dimension. This represents a branch that leans towards semantic compression. This indicates a preference for high-fidelity branches.

[0066] Furthermore, in S3, the dynamic switching algorithm specifically includes:

[0067] S33. Setting Decision Thresholds and Branch Selection ,when When the system automatically selects the GCM-SeC model to perform cross-modal semantic compression and generative reconstruction, i.e., the text compression-generative reconstruction path, it improves bandwidth efficiency while ensuring semantic consistency; when When the time comes, switch to the SwinJSCC model, i.e., the high-fidelity image direct transmission path, to perform end-to-end high-fidelity image feature encoding and transmission;

[0068] S34. Weight Parameter Tuning and Adaptive Configuration , The initial parameter values ​​are preset based on the on-site environment or task type, and are subsequently adaptively fine-tuned according to the relative importance of task requirements.

[0069] Furthermore, in S4, the dual-state satellite-to-ground link simulation model specifically includes:

[0070] S41. Dynamic State Division of Satellite-Ground Channel: The simulation model divides the satellite-ground transmission link into two states: good state and bad state, which correspond to different channel interference environments.

[0071] S42. Good-state modeling: Under good conditions, the channel model includes the main path signal and multipath scattering components. The main path signal gain is modeled based on the Rician factor K in the Rician distribution. Multipath scattering and small-scale fading are modeled using the Rayleigh distribution, which comprehensively reflects the characteristics of large-scale and small-scale fading.

[0072] S43. Modeling under adverse conditions: Under adverse conditions, the main path of the channel is significantly affected by shadowing fading. Rayleigh distribution is used to simulate multipath fading, and the shadowing fading factor follows a log-normal distribution.

[0073] Furthermore, in S4, the dual-state satellite-to-ground link simulation model specifically includes:

[0074] S44. The state transition and sampling mechanism adopts a probabilistic model. The parameters are set by the system or scenario requirements to determine the channel state for each sampling, enabling dynamic switching and sample generation under good / bad conditions;

[0075] S45. The parameter generation and transformation model initializes and samples the real part of the channel for each state. With the imaginary part The time-domain channel response is obtained. And use Fast Fourier Transform to convert it into the frequency domain channel response. ;

[0076] S46. The Doppler frequency shift simulation generates the channel frequency domain response, which is further superimposed with motion-related Doppler phase rotation.

[0077] S47. Integrated Interference Simulation and Parameter Customization: Based on the application scenario, the Rician factor, shadow fading parameters, and Doppler frequency offset simulation parameters are customized to achieve multi-factor joint interference modeling adapted to typical 6G satellite-to-ground communication scenarios.

[0078] The present invention also relates to a cross-modal generative satellite-to-ground semantic communication system with ultra-high compression ratio, comprising a computer module that applies the above-described method.

[0079] Beneficial effects

[0080] This invention proposes a cross-modal generative semantic communication method for satellite-to-ground networks (GCM-SeC) with ultra-high compression ratio. In single-user scenarios, this method efficiently compresses the original image information into a compact text semantic representation through a cross-modal semantic compression module. It adopts an end-to-end generative model to ensure high semantic consistency and robust reconstruction, achieving a significant reduction in data volume compression and low-latency semantic transmission, effectively reducing the bandwidth burden of satellite-to-ground networks.

[0081] This invention also proposes a GCM-SeC-NOMA model, which deeply integrates non-orthogonal multiple access technology (NOMA) with cross-modal semantic communication GCM-SeC, significantly improving the system spectrum utilization rate. It can effectively meet the needs of multi-user concurrent access and resource sharing in the satellite-ground communication environment, and realize efficient access for large-scale multi-users.

[0082] To address the diverse mission requirements and variable channel issues in satellite-to-ground scenarios, this invention innovatively introduces a dynamic switching algorithm. This algorithm can intelligently select between an ultra-high compression semantic generation path and a high-fidelity image direct transmission path based on the channel signal-to-noise ratio and the specific mission type. This achieves an adaptive dynamic balance between bandwidth, semantic consistency, and image fidelity, meeting the differentiated transmission requirements of various complex application scenarios such as emergency response and long-distance transmission.

[0083] Furthermore, this invention establishes a dynamic dual-state satellite-to-ground link integrated simulation model, comprehensively considering typical satellite-to-ground interference factors such as shadowing, multipath propagation, and Doppler shift, providing strong support for algorithm performance verification in complex environments. Extensive simulation results show that this invention can compress image transmission data to 0.085% or less of the original volume, achieve a semantic consistency BLEU-4 score greater than 0.92, and improve semantic transmission accuracy to over 0.94, representing an improvement of 3.3%–12.9% compared to the existing LAM-MSC scheme. Moreover, it can linearly improve spectral efficiency with the number of users. Attached Figure Description

[0084] Figure 1 This is a diagram of the overall architecture of GCM-SeC in this invention;

[0085] Figure 2 This is an architecture diagram of the cross-modal semantic compression module in this invention;

[0086] Figure 3 This is an architecture diagram of the cross-modal generation and reconstruction module in this invention;

[0087] Figure 4 This is a diagram of the overall architecture of GCM-SeC-NOMA in this invention;

[0088] Figure 5 This is a flowchart of the dynamic switching algorithm in this invention;

[0089] Figure 6a This is a comparison chart of the BLEU-4 scores of the GCM-SeC system in this invention under different channels;

[0090] Figure 6b This is a comparison chart of CLIP scores for the GCM-SeC system in this invention under different channels;

[0091] Figure 6cThis is a comparison chart of BERT cosine similarity of the GCM-SeC system in this invention under different channels;

[0092] Figure 6d This is a comparison chart of the transmission accuracy of the GCM-SeC system in this invention under different channels;

[0093] Figure 7 This is a schematic diagram comparing the transmission accuracy of GCM-SeC and LAM-MSC in the VOC2012 dataset of this invention;

[0094] Figure 8 This is a schematic diagram illustrating the visualization effect of GCM-SeC in the satellite-to-ground channel when the signal-to-noise ratio is 5 in this invention;

[0095] Figure 9 This is a schematic diagram illustrating an embodiment of the dynamic switching algorithm in this invention;

[0096] Figure 10a This is a comparison chart of the BLEU-4 scores of two users under the satellite-to-ground channel in this invention, specifically for GCM-SeC-NOMA and GCM-SeC-OMA.

[0097] Figure 10b This is a comparison chart of CLIP scores for two users under GCM-SeC-NOMA and GCM-SeC-OMA in the satellite-to-ground channel in this invention;

[0098] Figure 10c This is a comparison chart of BERT cosine similarity between two users in GCM-SeC-NOMA and GCM-SeC-OMA under satellite-to-ground channels in this invention;

[0099] Figure 10d This is a comparison chart of the transmission accuracy of two users under the satellite-to-ground channel in GCM-SeC-NOMA and GCM-SeC-OMA in this invention;

[0100] Figure 11 This is a schematic diagram comparing the spectral efficiency of different numbers of users in this invention. Detailed Implementation

[0101] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0102] The present invention provides a cross-modal generative satellite-to-ground network semantic communication method with ultra-high compression ratio, which mainly includes the following steps:

[0103] S1. Establish a generative cross-modal satellite-to-ground network semantic communication system: For single-user scenarios, the image data to be transmitted is transformed into a compact text semantic representation in the cross-modal semantic compression module, and then transmitted through the satellite-to-ground channel module. At the receiving end, the cross-modal generative reconstruction module uses a text-driven end-to-end generative model to reconstruct the image, achieving high semantic consistency, efficient compression, and robust communication.

[0104] S2. Establish a semantic communication system model for satellite-to-ground network integrating NOMA technology: For multi-user scenarios, under the NOMA mechanism, the signals of multiple users are superimposed in the power domain, and serial interference cancellation technology is used at the receiving end to separate the signals of each user.

[0105] S3. Establish a dynamic switching algorithm: Based on the real-time signal-to-noise ratio of the channel and the requirements of the task type, adaptively select the text compression-generative reconstruction path or the high-fidelity image direct transmission path to achieve a dynamic balance between bandwidth efficiency, semantic consistency and image fidelity.

[0106] S4. Establish a dual-state satellite-to-ground link simulation model: Taking into account the effects of multipath, shadowing and Doppler factors in satellite-to-ground communication, the satellite-to-ground transmission link is divided into two states: good and bad. The channel characteristics are modeled using Rician and Rayleigh distributions, and Doppler frequency shift is simulated by Doppler phase rotation to complete the satellite-to-ground channel modeling and simulation.

[0107] The system model applied in this invention is as follows:

[0108] GCM-SeC model architecture

[0109] To address the challenges of spectrum constraints, poor channel conditions, and diverse missions in space-ground converged networks, this invention proposes GCM-SeC. Its system architecture is as follows: Figure 1 As shown, the whole consists of three main components: a cross-modal semantic compression module, a satellite-to-ground channel model, and a cross-modal generative reconstruction module.

[0110] ① Cross-modal semantic compression module: This module is deployed at the transmitting end. Its main function is to efficiently compress the original image S into a cross-modal text semantic representation X, thereby reducing the transmission load and enhancing channel robustness. For image S, the image semantic feature extraction module first transforms it into the corresponding text semantic representation T. Then, the joint source-channel encoder module transforms the text representation T into the vector X that the channel needs to transmit, i.e.

[0111]

[0112] in, θ represents the joint source-channel coding, θ represents the parameters of the neural network, and SNR is the signal-to-noise ratio.

[0113] ② Satellite-to-ground channel transmission module: Considering the special channel characteristics of satellite-to-ground links, this invention performs unified simulation modeling of satellite-to-ground channels, comprehensively considering factors such as path loss, Doppler shift, and shadowing fading.

[0114]

[0115] in, For satellite-to-ground channel gain, It is additive Gaussian noise.

[0116] ③ Cross-modal generative reconstruction module: This module is deployed at the receiving end and aims to reconstruct an image from the received text semantics Y. This ensures low latency and high reliability transmission of basic semantics. First, a joint source-channel decoder converts the signal Y received from the channel into the corresponding textual semantic representation. ,Right now

[0117]

[0118] in, Indicates joint source-channel decoding, These represent the parameters of the neural network. Then, the image generative reconstruction module converts them into a reconstructed image. .

[0119] In summary, GCM-SeC constructs a complete cross-modal semantic communication link, from image semantic extraction, text representation generation, channel transmission to text-driven image reconstruction. Its architecture possesses strong scalability and task adaptability. Further development could incorporate NOMA mechanisms and dynamic switching algorithms to adapt to the communication needs of satellite-to-ground networks under complex conditions such as low signal-to-noise ratio, high dynamism, and limited spectrum.

[0120] Dynamic dual-state satellite-to-ground link simulation model

[0121] In satellite mobile communication, the relative motion between the transmitter and receiver causes the transmission path to constantly change. Throughout the communication process, numerous interferences exist, including shadowing, multipath propagation, and Doppler shift. Therefore, when modeling the satellite-to-ground link, the impact of these interferences on the channel transmission environment needs to be comprehensively considered.

[0122] Currently, common satellite channel models are shown in Table 1. These three types of models can fully simulate the performance of different satellites and cover many interferences in satellite-to-ground network communication, such as multipath propagation, shadowing fading, and Doppler shift, thereby simulating the real satellite communication environment and providing a foundation for related research in satellite-to-ground networks. However, each of the three models has its advantages and disadvantages, but none of them fully considers the various interferences that may occur during transmission. Therefore, in order to make the satellite-to-ground link more realistic, this invention considers the above three types of interference and constructs a unified dual-state satellite-to-ground link simulation model to fully simulate the harsh environment of satellite-to-ground communication.

[0123] Table 1 Comparison of different satellite channel models

[0124] characteristic C.Loo Corazza Lutz Shadow decay √ √ √ Multipath propagation √ √ √ Doppler shift × √ / × × Implementation complexity lower high medium

[0125] The specific process is shown in Algorithm 1. The entire channel modeling process is divided into two states: a good state and a bad state. In the good state, the channel uses the Ricean channel model to simulate the main path signal and the scattered path signal. The main path signal contains a strong direct path, and its gain is determined by the Ricean factor, while the scattered path signal follows a Rayleigh distribution. This method can effectively simulate the large-scale and small-scale fading characteristics in the channel. In the bad state, the channel uses the shadowed fading Rayleigh channel model, where the path gain is affected by shadowed fading, the shadowed fading factor follows a log-normal distribution, and small-scale fading is simulated by a Rayleigh distribution.

[0126] Algorithm 1 Satellite-to-Ground Link Simulation Algorithm Input: Number of samples num_sample, good-state probability P_good, Rice factor K, shadow fading standard deviation sigma_shadow_dB Output: h_fft (frequency domain CSI samples) 1 (1) Initialize parameters: 2 Calculate model-related parameters and initialize h_I and h_Q 3 (2) Generate num_sample CSI samples in a loop: 4 For i = 0 tonum_sample-1 do: 5 Determine the channel state (based on P_good) 6 If state == 'good' then: 7 Select good state parameters 8 Else: 9 Select bad state parameters and generate shadow factor shadow 10 For each path delay d do: 11 If state == 'good' then: 12 Generate h_I and h_Q according to Rice distribution (distinguish between principal path and scattering path) 13 Else: 14 Generate h_I and h_Q according to shadow Rayleigh distribution 15 (3) Convert time domain to frequency domain: 16 Convert h_I and h_Q to complex tensors h 17 Perform FFT transformation on h to obtain h_fft 18 (4) Apply frequency-related Doppler phase: 19 Calculate the frequency-related phase 20 Apply the phase to h_fft Returns: h_fft

[0127] In addition, the channel also considers the effect of Doppler frequency shift, which is simulated by applying phase rotation on the time delay path. Table 2 shows the specific satellite-to-ground channel parameter settings.

[0128] Table 2 Simulation Modeling Parameter Settings

[0129] parameter Parameter value Probability of good state 0.6 Rice factor 4 Shadow fading standard deviation 3 bandwidth 10Mbps

[0130] The main methods applied in the cross-modal generative satellite-to-ground network semantic communication with ultra-high compression ratio of the present invention are as follows.

[0131] Cross-modal semantic compression module

[0132] The cross-modal semantic compression module, located at the transmitter of the GCM-SeC system, is a key component for achieving low-resource image transmission. Its core objective is to transform the high-dimensional visual semantic features inherent in the original image into a compact, low-transmission-cost natural language description, significantly compressing data volume and improving the robustness of the satellite-to-ground channel. The specific architecture is as follows: Figure 2 As shown.

[0133] First, in the feature extraction stage, this invention adopts the latest BLIP-2 (Bootstrapped LanguageImage Pretraining v2) architecture, combining a frozen visual encoder, a frozen large language model (LLM), and a trainable intermediate module—the query transformer (Q-Former)—to construct an efficient vision-language bridging path. Specifically, assuming the transmitted image is... , and These represent the height and width of the image, respectively. After passing through the VIT image encoder, a set of image feature vectors will be generated:

[0134]

[0135] Where N is the number of blocks in the image and d is the feature dimension of the block.

[0136] Next, the system introduces M learnable query vectors. As input to the Q-Former, it interacts with itself using a self-attention mechanism and is fused with image features V through cross-attention to extract the semantic information most crucial for text generation:

[0137]

[0138] In BLIP-2's two-stage pre-training strategy, Q-Former, in the first stage, jointly optimizes image-text comparison, image caption generation, and matching tasks, learning how to extract text-relevant representations from visual features. The second stage then passes the Q-Former's output through a projection layer. Input space required for projection to LLM:

[0139]

[0140] This visual guidance prompt is prefixed to the text sequence and input into the frozen LLM, guiding it to generate a natural language description corresponding to the image. ,in Indicates the i-th word, For the vocabulary list:

[0141]

[0142] After generating the text, this invention further uses BERT (Bidirectional Encoder Representations from Transformers) to perform semantic embedding encoding on the natural language sequence. This model, based on the Transformer architecture, captures semantic information in the text through bidirectional context, encoding the text sequence T into a context-sensitive vector representation. ,Right now .

[0143] Through the above process, GCM-SeC achieves semantic cross-modal compression from image to text, significantly reducing transmission load while ensuring semantic consistency, and greatly enhancing the system's practicality and robustness in complex satellite-to-ground channels.

[0144] Cross-modal generative reconstruction module

[0145] The cross-modal generative reconstruction module's task is to restore the compressed semantic signal received from the satellite-to-ground link into image content with semantic consistency and visual fidelity. This module comprises two sub-modules: a joint decoding module and a text-driven image generation module, which respectively handle the text restoration and image reconstruction processes. The specific architecture is as follows: Figure 3 As shown.

[0146] First, the system maps the received channel output Y into a natural language description using a text inverse encoder, Transformer. This process is autoregressive:

[0147]

[0148] The generated text The core semantic components of the image are preserved, and can be used for subsequent image generation.

[0149] In the image reconstruction stage, this invention employs the text-driven diffusion-based image generation model StableDiffusion. This model learns the image distribution under textual conditions in the latent space and achieves high-quality image generation through an anti-diffusion process. Specifically, firstly, The input text encoder obtains the condition vector. Then Gaussian noise T-step anti-diffusion is performed:

[0150]

[0151] During the generation process, a consistent random seed and a fixed inference step size are used to control the diversity and stability of the images.

[0152] Meanwhile, to achieve end-to-end learning from image to text and back to image, GCM-SeC introduces various loss functions during the training phase, including semantic consistency scores. Mean square error loss Cross-entropy loss The specific optimization objective is to minimize the following reconstruction loss:

[0153]

[0154] Where λ is the corresponding weight. The specific end-to-end algorithm is shown in Figure 2.

[0155] Algorithm 2 GCM-SeC Training Process Input: Image dataset S, learning rate Training epochs N, output: trained model parameters Initialization parameters For epochs = 1 to N, do: For each image Dataset do: / / Semantic consistency loss / / Mean squared error loss / / Cross-entropy loss Backpropagation and parameter update: End for Return: Parameters

[0156] GCM-SeC-NOMA Multi-User Access Mechanism

[0157] To improve the utilization efficiency of spectrum resources in the space-to-ground network and meet the needs of simultaneous access by a large number of terminal devices, the GCM-SeC system can selectively introduce a non-orthogonal multiple access (NOMA) mechanism, integrating it with the semantic communication architecture to achieve multi-user concurrent transmission. The specific architecture is as follows: Figure 4 As shown.

[0158] Specifically, each user's semantic encoding representation After power weighting, the signals are superimposed in the power domain to form a composite signal:

[0159]

[0160] in, The power factor allocated to the i-th user satisfies At the receiving end, the system employs Successive Interference Cancellation (SIC) technology to decode the received signal Y. Taking a two-user scenario as an example, the receiving end prioritizes decoding the signal from the user with higher power. Other signals are treated as interference. After decoding, they are reconstructed and removed from the superimposed signal, thereby recovering the lower-power user signal. :

[0161]

[0162] In this way, subsequent users will no longer be affected by prior signals, and the overall anti-interference capability of the system will be significantly enhanced. This mechanism effectively realizes spectrum resource sharing, supports multiple users to access and transmit simultaneously without increasing bandwidth, and significantly improves system capacity and semantic communication concurrency capability.

[0163] Dynamic switching algorithm

[0164] The GCM-SeC system effectively reduces data transmission volume and improves robustness, but sacrifices image detail, making it suitable for resource-constrained, low-latency, high-interference, or urgent task scenarios. Existing semantic communication systems based on the Swin Transformer (hereinafter referred to as SwinJSCC) can learn hierarchical semantic features and image details, significantly improving image detail reconstruction, but with high data transmission volume, making them suitable for tasks with good channel conditions, sufficient bandwidth, and high requirements for image restoration accuracy. These two approaches are contradictory yet complementary.

[0165] In satellite-to-ground communication scenarios, different tasks have different requirements, and therefore different systems are applicable. Therefore, in order to better complete various tasks in satellite-to-ground communication scenarios, it is urgent to design a dynamic switching algorithm that can intelligently activate the most suitable communication mode according to the channel status and task requirements to achieve the optimal transmission strategy.

[0166] ① Channel and mission status parameter input

[0167] The real-time channel signal-to-noise ratio (SNR) of the received input and the task requirement vector Each dimension of the task vector These correspond to specific requirements in satellite-to-ground scenarios, such as semantic accuracy, latency sensitivity, data volume constraints, and energy consumption tolerance.

[0168] ② Branch scoring function construction: Based on the current channel state and task vector, the SNR range is set. and A comprehensive scoring function is established by normalizing the signal-to-noise ratio and weighting the task score:

[0169]

[0170] parameter , Used to adjust the relative weights of channel evaluation and mission requirements in branch decisions. The importance weights representing the requirements of each dimension can be flexibly configured according to the task scenario. (Task dimension) This represents the semantic compression-oriented branch (GCM-SeC). This represents a high-fidelity branch (SwinJSCC);

[0171] ③ Decision Threshold and Branch Selection: Setting Decision Thresholds ,when The system automatically selects the GCM-SeC path to perform cross-modal semantic compression and generative reconstruction, which is suitable for scenarios with limited bandwidth, time sensitivity and strong interference; otherwise, it switches to the SwingJSCC path to perform end-to-end high-fidelity image feature encoding and direct transmission, which meets the requirements of high bandwidth or high-quality channel environments with high requirements for image details.

[0172] ④ Weight parameter tuning and adaptive configuration , The initial parameter values ​​are preset according to the field environment or task type, and can be adaptively fine-tuned according to the relative importance of task requirements to ensure that the switching algorithm can take into account both channel adaptability and multi-task coordination in actual operation.

[0173] The relevant dynamic switching strategy process is as follows: Figure 5 As shown, the pseudocode is as shown in Algorithm 3:

[0174] Algorithm 3: Dynamic Switching Decision Algorithm Input: Current signal-to-noise ratio (SNR), task type label threshold Weighting coefficient Output: Branch selection result: selected_branch Set the signal-to-noise ratio range SNR_min and SNR_max, and set the scoring threshold. Set scoring weights , if S < then selected_branch ← "GCM-SeC"Else selected_branch ← "SwinJSCC"end if Returns: selected_branch

[0175] Effect verification

[0176] In one embodiment of this invention, the relevant parameters are shown in Table 3. Meanwhile, to comprehensively evaluate the performance of the GCM-SeC system in a satellite-ground fusion network, this invention uses the MS-COCO dataset and the VOC2012 dataset as image sources. Based on different SNR (signal-to-noise ratio) conditions and task requirements, the advantages and disadvantages of the proposed method compared with traditional methods are evaluated. The experimental platform uses a single NVIDIA RTX 4090 GPU for training, and the PyTorch framework is used for model training and inference.

[0177] To simulate the harsh channel environments commonly encountered in satellite-to-ground communication, this invention employs the proposed dynamic dual-state satellite-to-ground link simulation model. In the experiments, six different signal-to-noise ratio (SNR) values ​​were set: 0dB, 5dB, 10dB, 15dB, 20dB, and 25dB, to examine the impact of channel conditions on communication performance. Furthermore, to simulate spectrum resource usage in a multi-user environment, NOMA technology was used to simulate a scenario where multiple users share the same spectrum.

[0178] To comprehensively evaluate the performance of the GCM-SeC system, this invention employs a variety of evaluation metrics. These metrics can measure the system's performance in image compression, semantic transmission, image reconstruction, and multi-user concurrency from different perspectives.

[0179] BLEU-4 score: Used to measure the similarity between the reconstructed text and the original text to assess the accuracy of semantic reconstruction. A higher score indicates that the generated text is closer to the true description. The calculation formula is as follows:

[0180]

[0181] in, It is n-gram precision. It is an n-gram weight, and BP is a penalty factor used to reduce the situation where the generated text is too short.

[0182] CLIP score: The CLIP model is used to calculate the similarity between an image and text, serving as an evaluation metric for image-text semantic consistency. The CLIP score is calculated based on the latent spaces of the image and text. The degree of matching between the text description and the generated image is obtained by calculating the cosine similarity between the image and text embeddings. The formula is as follows:

[0183]

[0184] in, and Let denot and represent the embedding vectors of the image and text, respectively. cos represents the cosine similarity. The closer the result is to 1, the more similar the image and text are.

[0185] BERT Cosine Similarity: BERT cosine similarity is used to calculate the semantic similarity between the original text and the generated image description. By calculating the cosine similarity between the BERT feature vectors transformed from the original text and the generated description, this paper quantifies the similarity between the two. The formula for calculating BERT cosine similarity is:

[0186]

[0187] in, and These are the BERT embedding vectors of the original text and the generated description, respectively. The closer the cosine similarity is to 1, the higher the similarity between the two in the semantic space.

[0188] Transmission accuracy

[0189] A key metric for evaluating whether the reconstructed text transmitted by GCM-SeC matches the original text under given conditions is its accuracy in information transmission. It assesses the model's accuracy by calculating the BERT cosine similarity between the generated image description and the original text. Specifically, transmission accuracy is calculated by setting a similarity threshold (0.8 in this paper) and determining the proportion of successful transmissions exceeding this threshold. The formula is as follows:

[0190]

[0191] Where N is the total number of samples, It is an indicator function that indicates that when the cosine similarity is greater than 0.8, it is recorded as 1, otherwise it is 0.

[0192] Image reconstruction quality (PSNR vs. SSIM)

[0193] Used to evaluate image fidelity, a higher PSNR indicates a better image quality, and a closer SSIM is to 1.

[0194] System spectrum utilization

[0195] This study aims to evaluate the spectrum resource utilization efficiency of GCM-SeC-NOMA in multi-user environments, particularly its improvement under the NOMA mechanism. and The single-user spectrum efficiency can be derived as follows: .

[0196] Table 3 Basic parameter settings for GCM-SeC method simulation

[0197] parameter Value parameter Value Image size 224×224 Batch size 8 / 16 Learning rate Hidden layer dimensions 1024 Signal-to-noise ratio [0,25] Encoding Dimensions 96 Stop early and be patient 6 Number of attention heads 8 Weight 0.1 Transformer layer number 2 Weight 0.25 Validation set partition ratio 0.2 Weight 0.65 Coding efficiency 1.0 threshold 0.75 Power allocation between two users 0.6 / 0.4

[0198] GCM-SeC System Performance

[0199] Figure 6 illustrates the performance of GCM-SeC across different channels, including AWGN, Rayleigh, and STL. In this experiment, the input image was first converted into a textual semantic representation, from which 512×128 features were extracted and compressed into a 128-dimensional vector using a linear layer. This vector was then transmitted through the channel, and the receiver reconstructed the image using a linear layer and a decoder. Ultimately, the data for all three channels was compressed to 0.085% of the original image.

[0200] As shown in Figure 6(a), the BLEU-4 score consistently exceeds 0.9 under various channel conditions, indicating that the model possesses high accuracy in text reconstruction and can achieve reliable semantic transmission. During training, GCM-SeC uses SNR as an input parameter and calculates the overall loss to adapt to different channel conditions. Although this causes slight fluctuations in the BLEU-4 index, the model still converges to optimal performance overall.

[0201] In Figure 6(b), the CLIP score remained around 25 across all channels. This is because the compressed extracted text only retained the core semantic content of the image to minimize transmission overhead, thus failing to cover all image details. However, Figure 6(c) shows that the cosine similarity between the original and reconstructed images consistently exceeded 0.9, indicating that the reconstructed image successfully preserved the core semantic content. According to the semantic correctness criteria proposed in existing literature, semantic transmission can be considered successful when the cosine similarity is higher than 0.8. Based on this criterion, Figure 6(d) shows that the transmission accuracy exceeded 0.9 across all channels.

[0202] Figure 7The performance of GCM-SeC and LAM-MSC was compared under different channels. Although LAM-MSC uses a lower decision threshold (0.6), GCM-SeC still achieved higher transmission accuracy. In addition, as shown in Table 4, GCM-SeC has a significantly lower memory usage during the inference phase than LAM-MSC, and thanks to the adaptive SNR mechanism, it only needs to store one set of encoding and decoding parameters, making it more suitable for deployment on actual devices.

[0203] Table 4. Image comparison of GCM-SEC, SwinJSCC, and LAM-MSC

[0204] index GCM-SeC SwinJSCC LAM-MSC Data transfer volume (FP) 128 1204224 128 Transmission time (seconds) 0.0004096 3.8535 0.0004096 Processing time (seconds) 4.43 3.82 9.94 Total end-to-end time (seconds) 4.4304096 7.6735 9.9404096 GPU memory usage (MiB) 10702 1830 23962

[0205] Figure 8 A visualization example of the satellite-to-ground link (SNR of 5) is provided, demonstrating that GCM-SeC can still achieve text reconstruction with a high degree of semantic consistency with the original input image under extreme channel conditions, verifying its robust protection of core semantic information. It is worth noting that the last column in the figure is based on real image evaluation, with a BERT cosine similarity of 0.9727, further proving that this method achieves efficient image semantic transmission.

[0206] Overall, GCM-SeC maintains excellent performance across multiple channel types, including Gaussian, Rayleigh, and STL. Particularly in the STL environment, its performance is nearly on par with AWGN, demonstrating the model's strong robustness and adaptability. Furthermore, in visualization experiments under STL, the reconstructed text and the original image semantics are almost identical, with a BERT cosine similarity of 0.9727, verifying that the model can still achieve accurate basic semantic transmission under extreme channel conditions.

[0207] Dynamic switching method effect verification

[0208] In GCM-SeC, the input image first undergoes cross-modal feature extraction and semantic text encoding to generate an intermediate feature vector B×512×F, where B represents the batch size and F is the feature dimension. Before channel transmission, these features are mapped to a low-dimensional space K through linear projection. When K is small, the required data transmission, transmission latency, and bandwidth usage are all reduced. In contrast, SwinJSCC employs deep joint source-channel coding, encoding the image into a feature vector of B×3136×F.

[0209] Table 5 shows the performance comparison of various methods in the STL environment under different feature dimensions and signal-to-noise ratios. In the GCM-SeC section, the two sets of numbers in the table represent F and K, respectively. As F gradually decreases, both the average BLEU-4 score and CLIP score decrease, because a smaller F reduces the number of extractable text features. As can be seen from the table, F=192 and F=96 have similar performance, significantly reducing bandwidth and data volume burden while maintaining a certain level of performance; therefore, F=96 is the optimal choice. When F is fixed, decreasing K also leads to a slight decrease in the BLEU-4 score. Furthermore, due to the good robustness of the BLEU metric, other metrics (such as CLIP and BERT cosine similarity) also show relative balance and stability under different K values. The results show that GCM-SeC can work stably and robustly throughout the entire signal-to-noise ratio range, and greatly reduces transmission overhead. When F=96 and K=128, the data is compressed to 0.085% of the original image. It should be noted that only basic semantic information is transmitted at this time, and pixel-level details cannot be recovered. Therefore, GCM-SeC is particularly suitable for scenarios with limited bandwidth or severe interference.

[0210] Table 5 Comparison of STL performance under different compression ratios and SNR conditions

[0211]

[0212]

[0213]

[0214] For SwinJSCC, when F=384, the required data transmission is 8 times that of the original image, and even at F=192, it is still 4 times, significantly higher than the bandwidth and data transmission required by GCM-SeC. As F or SNR decreases, its PSNR and SSIM indicators decrease synchronously. Furthermore, Table 4 shows that SwinJSCC is far superior to GCM-SeC in both data transmission volume and latency. Although SwinJSCC has poor robustness and latency performance, it can effectively preserve rich visual details in high-resolution, high-fidelity applications. Both schemes have their advantages; therefore, the appropriate transmission strategy should be flexibly selected based on specific task requirements.

[0215] Figure 9 The effectiveness of the proposed dynamic switching algorithm was verified by selecting two representative tasks: emergency disaster relief and remote sensing reconnaissance. The former only requires rapid transmission of core semantic information and does not have strict requirements for high-precision reconstruction; the latter requires high-fidelity restoration of image details for detailed interpretation. Therefore, the core metrics are latency and data volume. The feature vectors for the two tasks are [0, 0] and [1, 1], respectively. and Both are set to 0.5, and the environment and task weights ω1 and ω2 are also set to 0.5. When the channel conditions are good and there are no requirements for data volume or latency, SwinJSCC is selected first, so the threshold τ is set to 0.75.

[0216] In this setup, using GCM-SeC alone can achieve high semantic consistency (BERT score of 0.979) in remote sensing scenarios, but it struggles to recover pixel-level details, potentially leading to the loss of some critical information. In emergency scenarios, using SwinJSCC alone results in approximately eight times the data volume of the original image, severely impacting transmission timeliness under limited bandwidth, and its performance degrades significantly under low SNR conditions. In contrast, the proposed dynamic switching algorithm (the part within the black dashed box in the figure) adaptively selects the optimal strategy based on channel and task type, combining the advantages of both branches to fully meet the needs of diverse tasks.

[0217] Evaluation of GCM-SeC-NOMA and Spectral Efficiency

[0218] Figure 10 evaluates the transmission performance of GCM-SeC-NOMA and GCM-SeC-OMA during simultaneous dual-user communication over a satellite-to-ground channel. Firstly, although the performance metrics of GCM-SeC-NOMA and GCM-SeC-OMA are quite similar, from... Figure 11 It can be seen that NOMA has higher spectral efficiency than OMA (such as TDMA or FDMA). Other OMA schemes (such as CDMA or OFDMA) still have lower spectral efficiency than NOMA due to limitations in orthogonal resource block allocation. Therefore, this paper adopts NOMA to improve spectrum utilization and solve the problem of multi-user synchronous access. Secondly, under the above conditions, each user only needs to transmit 0.085% of the original image data. Regardless of channel quality, the BERT cosine similarity between the original and reconstructed images of the two users is higher than 0.9, and the semantic transmission accuracy also exceeds 0.9, indicating that GCM-SeC-NOMA still has performance stability and good scalability when the number of users increases. At the same time, the performance difference between the two users is minimal, indicating that when the power allocation is relatively balanced, all users can obtain ideal reconstruction quality.

[0219] Figure 11 The spectral efficiency of GCM-SeC and GCM-SeC-NOMA in multi-user scenarios was compared. The results show that spectral efficiency increases with increasing SNR. With NOMA, users share the power domain within the same time slot, achieving synchronous data transmission among multiple users. Therefore, the total system spectral efficiency equals the sum of the spectral efficiencies of all users. Based on this principle, the spectral efficiency of GCM-SeC-NOMA can increase proportionally with the number of users while maintaining stable performance for each user, effectively solving the challenge of large-scale concurrent access in satellite-to-ground communication systems.

[0220] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A cross-modal generative satellite-to-ground network semantic communication method with ultra-high compression ratio, characterized in that, Includes the following steps: S1. Establish a generative cross-modal satellite-ground network semantic communication system (GCM-SeC): For single-user scenarios, the image data to be transmitted is transformed into a compact text semantic representation in the cross-modal semantic compression module, and then transmitted through the satellite-ground channel module. At the receiving end, the cross-modal generative reconstruction module uses a text-driven end-to-end generative model to reconstruct the image, achieving high semantic consistency, efficient compression, and robust communication. S2. Establish a semantic communication system model for satellite-to-ground network integrating NOMA technology: For multi-user scenarios, under the NOMA mechanism, the signals of multiple users are superimposed in the power domain, and serial interference cancellation technology is used at the receiving end to separate the signals of each user. S3. Establish a dynamic switching algorithm: Based on the real-time signal-to-noise ratio of the channel and the requirements of the task type, adaptively select the text compression-generative reconstruction path or the high-fidelity image direct transmission path to achieve a dynamic balance between bandwidth efficiency, semantic consistency and image fidelity. S4. Establish a dual-state satellite-to-ground link simulation model: Taking into account the effects of multipath, shadowing and Doppler factors in satellite-to-ground communication, the satellite-to-ground transmission link is divided into two states: good and bad. The channel characteristics are modeled using Rician and Rayleigh distributions, and Doppler frequency shift is simulated by Doppler phase rotation to complete the satellite-to-ground channel modeling and simulation. In S1, the specific method of the generative cross-modal satellite-to-ground network semantic communication system includes: S11, Cross-modal semantic compression module In the feature extraction stage, the BLIP-2 architecture is adopted, combining a frozen visual encoder, a frozen large language model, and a trainable intermediate module—a query transformer—to construct a visual-language bridging path; the transmitted image is... , and These represent the height and width of the image, respectively. After passing through the ViT image encoder, a set of image feature vectors will be generated: Where N is the number of blocks in the image, and d is the feature dimension of the block; Introduce M learnable query vectors As input to the Q-Former, it interacts with itself using a self-attention mechanism and is fused with image features V through cross-attention to extract the semantic information most crucial for text generation: BLIP-2 includes a two-stage pre-training strategy. The first stage optimizes the Q-Former by jointly performing image-text comparison, image caption generation, and matching tasks, learning how to extract strongly correlated representations from visual features. The second stage passes the output of the Q-Former through a projection layer. Input space required for projection to LLM: Visual guidance cues are prefixed to the text sequence and input into the frozen LLM, guiding it to generate natural language descriptions corresponding to the images. ,in Indicates the i-th word, Vocabulary list: After text generation, BERT is used to semantically embed the natural language sequence. This model, based on the Transformer architecture, captures semantic information in the text through bidirectional context, encoding the text sequence T into a context-sensitive vector representation. ,Right now ; S12, Satellite-to-Ground Channel Transmission Module A unified simulation model was performed for the satellite-to-ground channel, taking into account path loss, Doppler shift, and shadowing fading factors. in, For satellite-to-ground channel gain, It is additive Gaussian noise; S13, Cross-modal generative reconstruction module First, the system maps the received channel output Y into a natural language description using a text inverse encoder, Transformer. The generation process is autoregressive: ; generated text The core semantic components of the image are preserved, and can be used for subsequent image generation.

2. The cross-modal generative satellite-to-ground network semantic communication method with ultra-high compression ratio according to claim 1, characterized in that, In S1, the specific method of the generative cross-modal satellite-to-ground network semantic communication system further includes: In the image reconstruction stage, a text-driven diffusion-based image generation model, Stable Diffusion, is employed. This model learns the image distribution under textual conditions in the latent space and achieves high-quality image generation through an anti-diffusion process: First, the text is... The input text encoder obtains the condition vector. Then Gaussian noise T-step anti-diffusion is performed: ； During the generation process, a consistent random seed and a fixed inference step size are used to control the diversity and stability of the images; Meanwhile, GCM-SeC introduces multiple loss functions during the training phase, including semantic consistency scores. Mean square error loss Cross-entropy loss The specific optimization objective is to minimize the following reconstruction loss: ;in , , The corresponding weights.

3. The cross-modal generative satellite-to-ground network semantic communication method with ultra-high compression ratio according to claim 1, characterized in that, In S2, the satellite-to-ground network semantic communication system model integrating NOMA technology specifically includes: S21, Cross-modal semantic compression module For each user, the channel is obtained using the cross-modal semantic compression module in GCM-SeC. For image S, First, the image semantic feature extraction module transforms the text semantic representation T into a corresponding text semantic representation. Then, the joint source-channel encoder module transforms the text representation T into a vector that the channel needs to transmit. ,Right now ； in, θ represents the joint source-channel coding, θ represents the parameters of the neural network, and SNR is the signal-to-noise ratio; S22, Power Superposition For each user The semantic encoding represents power weighting, which is then superimposed on the power domain to form a composite signal: ； in, The power factor allocated to the i-th user satisfies .

4. The cross-modal generative satellite-to-ground network semantic communication method with ultra-high compression ratio according to claim 1, characterized in that, In S2, the satellite-to-ground network semantic communication system model incorporating NOMA technology specifically also includes: S23, Channel Transmission The resulting composite signal The signal will be transmitted via a satellite-to-ground link. : ； in, For satellite-to-ground channel gain, It is additive Gaussian noise; S24, Serial interference cancellation On the receiver side, the system uses continuous interference to cancel out the interference on the received signal. Decoding is performed; in a dual-user scenario, the receiver first decodes the signal from the user with higher power. Decoding is performed, treating signals from other users as interference; after decoding, the signal is reconstructed and subtracted from the composite signal, thereby recovering the signal from the lower-power user. : ； S25, Cross-modal generative reconstruction module For the decoded signals of each user First, the signal received from the channel is obtained through a joint source-channel decoder. Transform into the corresponding text semantic representation ,Right now ； in, Indicates joint source-channel decoding, The parameters of the neural network are then used by the image generative reconstruction module to convert them into a reconstructed image. .

5. The cross-modal generative satellite-to-ground network semantic communication method with ultra-high compression ratio according to claim 4, characterized in that, In S3, the dynamic switching algorithm specifically includes: S31. Channel and mission status parameter input Real-time channel signal-to-noise ratio of received input and task requirement vector Each dimension of the task vector These correspond to specific demand indicators in the space-to-ground scenario; S32. Branch scoring function construction: Based on the current channel state and task vector, set the SNR range. and A comprehensive scoring function is established by normalizing the signal-to-noise ratio and weighting the task score: ； parameter , Used to adjust the relative weights of channel evaluation and mission requirements in branch decisions. The importance weights of each dimension of requirements can be flexibly configured according to the task scenario, within the task dimension. This represents a branch that leans towards semantic compression. This indicates a preference for high-fidelity branches.

6. The cross-modal generative satellite-to-ground network semantic communication method with ultra-high compression ratio according to claim 5, characterized in that, In S3, the dynamic switching algorithm specifically includes: S33. Setting Decision Thresholds and Branch Selection ,when When the system automatically selects the GCM-SeC model to perform cross-modal semantic compression and generative reconstruction, i.e., the text compression-generative reconstruction path, it improves bandwidth efficiency while ensuring semantic consistency; when When the time comes, switch to the SwinJSCC model, i.e., the high-fidelity image direct transmission path, to perform end-to-end high-fidelity image feature encoding and transmission; S34. Weight Parameter Tuning and Adaptive Configuration , The initial parameter values ​​are preset based on the on-site environment or task type, and are subsequently adaptively fine-tuned according to the relative importance of task requirements.

7. The cross-modal generative satellite-to-ground network semantic communication method with ultra-high compression ratio according to claim 1, characterized in that, In S4, the dual-state satellite-to-ground link simulation model specifically includes: S41. Dynamic State Division of Satellite-Ground Channel: The simulation model divides the satellite-ground transmission link into two states: good state and bad state, which correspond to different channel interference environments. S42. Good-state modeling: Under good conditions, the channel model includes the main path signal and multipath scattering components. The main path signal gain is modeled based on the Rician factor K in the Rician distribution. Multipath scattering and small-scale fading are modeled using the Rayleigh distribution, which comprehensively reflects the characteristics of large-scale and small-scale fading. S43. Modeling under adverse conditions: Under adverse conditions, the main path of the channel is significantly affected by shadow fading. The Rayleigh distribution is used to simulate multipath fading, and the shadow fading factor follows a log-normal distribution.

8. The cross-modal generative satellite-to-ground network semantic communication method with ultra-high compression ratio according to claim 7, characterized in that, In S4, the dual-state satellite-to-ground link simulation model specifically includes: S44. The state transition and sampling mechanism adopts a probabilistic model. The parameters are set by the system or scenario requirements to determine the channel state for each sampling, enabling dynamic switching and sample generation under good / bad conditions; S45. The parameter generation and transformation model initializes and samples the real part of the channel for each state. With the imaginary part The time-domain channel response is obtained. And use Fast Fourier Transform to convert it into the frequency domain channel response. ; S46. The Doppler frequency shift simulation generates the channel frequency domain response, which is further superimposed with motion-related Doppler phase rotation. S47. Integrated Interference Simulation and Parameter Customization: Based on the application scenario, the Rician factor, shadow fading parameters, and Doppler frequency offset simulation parameters are customized to achieve multi-factor joint interference modeling adapted to typical 6G satellite-to-ground communication scenarios.

9. A cross-modal generative satellite-to-ground network semantic communication system with ultra-high compression ratio, characterized in that, It includes a computer module that applies the method described in any one of claims 1-8.

Images