Semantic Communication for Multi-Satellite Massive MIMO Transmission: A Mixture of Cooperative Modes Framework

Abstract: This paper investigates semantic communications (SemComs) for multi-satellite cooperative massive multiple-input multiple-output (MIMO) transmission, where multiple massive-MIMO satellites jointly serve a common set of multi-antenna user terminals. For the first time, SemComs with image transmission task are integrated into satellite massive MIMO and multi-satellite cooperative transmission. For the two representative cooperative modes, namely coherent transmission (CT) and non-coherent transmission (NCT), we develop multi-satellite CT (MSCT) and multi-satellite NCT (MSNCT) SemCom frameworks, respectively. MSCT adopts a symmetric architecture, whereas MSNCT introduces transmitter-side stream allocation and a two-stage receiver design that combines per-stream semantic extraction with cross-stream semantic-interference exploitation. To instantiate MSCT, we further design a symmetric encoder and decoder network based on hybrid Swin-Transformer and lightweight bottleneck convolutional neural network (CNN) blocks, termed HSTC, where Swin Transformer provides scalable computation and the CNN branch improves performance and convergence. For MSNCT, a Transformer-based backbone is employed to support cross-stream interference exploitation through global attention. Building on these two frameworks, we propose a mixture of cooperative modes (MoCM) framework, in which a permutation-invariant network dynamically switches between MSCT and MSNCT using multi-satellite statistical channel state information, thereby balancing semantic performance and complexity. Simulation results under practical configurations demonstrate the performance gains of the proposed frameworks.

• The paper introduces a cooperative-modes framework that adapts coherent and non-coherent transmission modes for semantic communications in multi-satellite MIMO, significantly improving image reconstruction quality.
• It employs a hybrid Swin Transformer and CNN architecture for CT and a two-stage Transformer for NCT, achieving notable gains in PSNR and SSIM across varying compression and channel scenarios.
• The MoCM controller dynamically selects transmission modes based on statistical CSI, effectively balancing computational complexity and semantic fidelity in satellite networks.

Semantic Communication for Multi-Satellite Massive MIMO: Mixture of Cooperative Modes Framework

Introduction and Motivation

The presented work advances the field of semantic communications (SemComs) by addressing the integration of SemComs into multi-satellite massive MIMO cooperative transmission systems. With the proliferation of mega-constellations and inter-satellite links (ISLs), exploiting the spatial, power, and coverage resources of multiple satellites becomes essential for handling the demands of post-5G and 6G networks in terms of reliability, capacity, and continuity. However, satellite systems operate with power constraints and experience highly challenging channel dynamics (e.g., large Doppler, delay spread), which make instantaneous CSI acquisition both resource-intensive and, in practice, infeasible. This substantiates the emphasis on statistical CSI (sCSI)-based designs.

Prior research on SemCom has predominantly targeted single-satellite-to-terminal links, largely neglecting the unique system and channel properties that arise in coordinated multi-satellite massive MIMO, particularly the distinct cooperative modes: coherent transmission (CT) and non-coherent transmission (NCT). This paper provides the first systematic framework for SemComs in such cooperative satellite systems, addressing both the physical and semantic layers, with a focus on image transmission tasks.

System Model and Problem Formulation

The study models the downlink of a multi-satellite system, where $S$ satellites, each equipped with large-scale UPAs, jointly serve multi-antenna user terminals (UTs). Leveraging sCSI ($$\{\gamma_{s,k}, K_{s,k}, \Omega_{s,k}, \Sigma_{s,k}\}_{s}$$), satellites coordinate transmissions while minimizing channel estimation overhead.

In CT, satellites deliver a single, coherent data stream to the UT, obtaining a link-budget gain via constructive combination. NCT, by contrast, assigns different streams to different satellites, which are received in parallel by the UT, leading to increased spatial multiplexing but also cross-stream interference. Both cases utilize spatial beamforming at the satellite and combining at the UT, but yield fundamentally different trade-offs in terms of complexity and achievable semantic fidelity.

The objective is to develop SemCom frameworks that are tailored to the properties and constraints of CT and NCT, enable efficient semantic transmission under massive MIMO multi-satellite paradigms, and dynamically optimize the system’s operational mode through a mixture-of-cooperative-modes (MoCM) mechanism.

Proposed Semantic Communication Frameworks

Multi-Satellite CT SemCom (MSCT)

MSCT exploits the coherent superposition of a single data stream transmitted by all cooperating satellites. Parameter-shared semantic encoders, instantiated with a hybrid Swin Transformer and bottleneck CNN (HSTC) backbone, are deployed at each satellite, ensuring both local adaptation and architectural symmetry. The decoder at the UT constructs a geometric sCSI-driven joint receive beam and incorporates statistical channel and effective noise features as auxiliary decoder input, enhancing semantic reconstruction robustness.

Notably, the adoption of the HSTC architecture ensures scalable computation through efficient windowed self-attention, maintains local semantic structure via bottleneck CNN modules, and accelerates convergence. Complexity analysis demonstrates that the per-satellite burden is dominated by semantic encoding, while the receiver leverages patch-based hierarchical structures for efficient decoding.

Multi-Satellite NCT SemCom (MSNCT)

For NCT, content is partitioned—either within an image (e.g., width-wise splitting) or by data batch—across satellites, each transmitting a unique semantic stream. Parameter sharing among semantic encoders is maintained, but input content varies.

The receiver decouples processing: An initial stream-specific semantic extraction is performed using a Transformer backbone, incorporating context from the full received signal and per-stream sCSI, followed by a cross-stream fusion stage using global self-attention to exploit and mitigate inter-stream interference at the semantic token level. This two-stage receiver efficiently combines parallelization, scalability, and semantic interaction, outperforming naive independent-stream approaches.

Mixture of Cooperative Modes Framework (MoCM)

Given the context-dependent benefits of CT (SNR-dominant, receiver-simplicity) and NCT (spatial multiplexing, lower per-satellite complexity), the MoCM framework introduces an adaptive, permutation-invariant neural network controller for mode selection. The selector receives low-dimensional, sCSI-derived feature sets across satellites and, via a satellite-dimension Transformer with set-attention pooling, produces mode selection logits, ensuring invariance to satellite labeling.

Offline, the selector is trained using cross-entropy loss, with ground-truth labels generated by benchmarking CT and NCT performance under various metrics (e.g., PSNR, SSIM) and complexity constraints per sample. This architecture enables dynamic trade-offs between semantic reconstruction performance and computational complexity according to application requirements and channel conditions.

Numerical Results and Empirical Validation

Simulation scenarios follow LEO mega-constellation parameters with practical array and link configurations. Performance is evaluated across a range of transmit powers, receiver antenna counts, and compression ratios, with CIFAR-10 as the image source.

Key Findings

• Superior Semantic Fidelity: Both MSCT and MSNCT consistently outperform BPG-LDPC and conventional direct SemCom extensions on PSNR and SSIM, across all power regimes and compression settings. Gains are particularly pronounced under more aggressive compression.
• Mode-Dependent Superiority: MSCT dominates in limited-antenna, lower SNR, or low compression settings (reflecting SNR-importance), while MSNCT prevails when spatial multiplexing is favored (larger arrays, higher compression).
• MoCM Efficiency: The MoCM framework closely tracks the mode-wise optimum by adaptively switching modes, achieving a superior performance/complexity trade-off. The network’s permutation-invariant selector remains robust to the number and labeling of satellites, and its complexity overhead is negligible relative to semantic encoding.
• Semantic Interference Exploitation: Decoder attention visualizations demonstrate that MSNCT effectively leverages cross-stream token interactions, enhancing global semantic consistency and enabling constructive handling of inter-stream interference, particularly when compared to direct, non-cross-stream decoders.
• Architectural Backbones: Under CT, the HSTC backbone offers the best performance-complexity trade-off; under NCT, pure Transformer backbones benefit joint token-domain attention and semantic fusion.

Implications and Future Directions

By synthesizing SemComs with multi-satellite cooperative massive MIMO, this work establishes a foundational framework for next-generation semantic-aware satellite networks. Practically, the proposed mechanisms can significantly enhance spectral efficiency, enable bandwidth- and power-constrained semantic services over wide-area satellite links, and power intelligent, adaptive data delivery.

The theoretical implications extend to the integration of task-oriented, information-centric communications with multi-user MIMO and distributed system designs. The cross-domain exploitation of statistical CSI, permutation-invariant neural architectures, and deep semantic coding is expected to inform both further satellite network innovations and terrestrial ultra-dense deployments.

Future extensions can address:

• Embedding temporal and multi-modal data flows.
• On-board learning and adaptation for per-satellite encoder adjustments.
• Extending semantic interference exploitation to general multi-stream, multi-user contexts.
• Integrating federated semantic learning across hybrid non-terrestrial and terrestrial infrastructures.

Conclusion

This work rigorously defines and addresses task-oriented semantic communications for multi-satellite massive MIMO transmission via flexible, mode-adaptive frameworks. By aligning semantic encoders/decoders with specific cooperative transmission modes and introducing an efficient selection mechanism, the study achieves pronounced semantic and computational gains, validating the feasibility and benefits of deep SemCom integration in space-based networks. The proposed MoCM, MSCT, and MSNCT frameworks form a solid basis for future semantic-aware wireless and satellite communication systems.