Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

S-MDMA: Sensitivity-Aware Model Division Multiple Access for Satellite-Ground Semantic Communication

Hui Cao, Rui Meng, Shujun Han, Song Gao, Xiaodong Xu, Ping Zhang

TL;DR

This work tackles bandwidth-limited satellite-ground SemCom by proposing S-MDMA, a sensitivity-aware model division multiple access framework. It decomposes semantic features into shared and differential parts, ranks features by reconstruction sensitivity, and uses a Kronecker-based orthogonal embedding to enable interference-free multi-user transmission. A lightweight geometric-mean reconstruction loss ensures balanced performance across users with varying SNRs, and a bandwidth-pruning step preserves only the most informative features under constrained channels. Experiments on DLRS and NWPU VHR-10 demonstrate that S-MDMA consistently outperforms Deep JSCC, WITT, and MDMA in PSNR and SSIM across diverse SNRs and datasets, with ablation results confirming the value of the sensitivity sorting and orthogonal embedding components. The approach offers a scalable, on-board friendly solution for robust, high-fidelity satellite-ground SemCom in real-world settings.

Abstract

Satellite-ground semantic communication (SemCom) is expected to play a pivotal role in convergence of communication and AI (ComAI), particularly in enabling intelligent and efficient multi-user data transmission. However, the inherent bandwidth constraints and user interference in satellite-ground systems pose significant challenges to semantic fidelity and transmission robustness. To address these issues, we propose a sensitivity-aware model division multiple access (S-MDMA) framework tailored for bandwidth-limited multi-user scenarios. The proposed framework first performs semantic extraction and merging based on the MDMA architecture to consolidate redundant information. To further improve transmission efficiency, a semantic sensitivity sorting algorithm is presented, which can selectively retain key semantic features. In addition, to mitigate inter-user interference, the framework incorporates orthogonal embedding of semantic features and introduces a multi-user reconstruction loss function to guide joint optimization. Experimental results on open-source datasets demonstrate that S-MDMA consistently outperforms existing methods, achieving robust and high-fidelity reconstruction across diverse signal-to-noise ratio (SNR) conditions and user configurations.

S-MDMA: Sensitivity-Aware Model Division Multiple Access for Satellite-Ground Semantic Communication

TL;DR

This work tackles bandwidth-limited satellite-ground SemCom by proposing S-MDMA, a sensitivity-aware model division multiple access framework. It decomposes semantic features into shared and differential parts, ranks features by reconstruction sensitivity, and uses a Kronecker-based orthogonal embedding to enable interference-free multi-user transmission. A lightweight geometric-mean reconstruction loss ensures balanced performance across users with varying SNRs, and a bandwidth-pruning step preserves only the most informative features under constrained channels. Experiments on DLRS and NWPU VHR-10 demonstrate that S-MDMA consistently outperforms Deep JSCC, WITT, and MDMA in PSNR and SSIM across diverse SNRs and datasets, with ablation results confirming the value of the sensitivity sorting and orthogonal embedding components. The approach offers a scalable, on-board friendly solution for robust, high-fidelity satellite-ground SemCom in real-world settings.

Abstract

Satellite-ground semantic communication (SemCom) is expected to play a pivotal role in convergence of communication and AI (ComAI), particularly in enabling intelligent and efficient multi-user data transmission. However, the inherent bandwidth constraints and user interference in satellite-ground systems pose significant challenges to semantic fidelity and transmission robustness. To address these issues, we propose a sensitivity-aware model division multiple access (S-MDMA) framework tailored for bandwidth-limited multi-user scenarios. The proposed framework first performs semantic extraction and merging based on the MDMA architecture to consolidate redundant information. To further improve transmission efficiency, a semantic sensitivity sorting algorithm is presented, which can selectively retain key semantic features. In addition, to mitigate inter-user interference, the framework incorporates orthogonal embedding of semantic features and introduces a multi-user reconstruction loss function to guide joint optimization. Experimental results on open-source datasets demonstrate that S-MDMA consistently outperforms existing methods, achieving robust and high-fidelity reconstruction across diverse signal-to-noise ratio (SNR) conditions and user configurations.
Paper Structure (29 sections, 1 theorem, 36 equations, 9 figures, 5 tables, 2 algorithms)

This paper contains 29 sections, 1 theorem, 36 equations, 9 figures, 5 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Works
  3. SemCom
  4. Multiple Access for SemCom
  5. Satellite-Ground SemCom
  6. Summary
  7. System Model
  8. Network Model
  9. Channel Model
  10. Overview of the Proposed Sensitivity-Aware Multiple Description Multiple Access (S-MDMA) Framework
  11. Proposed Semantic Similarity Fusion and Sort
  12. Similarity Semantic Feature Fusion
  13. Semantic Sensitivity Sorting and Cropping Algorithm
  14. Proposed Multi-user Data Transmission Module
  15. Semantic Feature Orthogonal Embedding
  16. ...and 14 more sections

Key Result

Lemma 1

${F}_{s-emb}$ and ${F}_{d-emb}$ are mutually orthogonal.

Figures (9)

  • Figure 1: The system model of multiuser satellite-ground SemCom, including LEO satellite, semantic transmission link, and ground terminals.
  • Figure 2: The architecture of the proposed S-MDMA scheme, where transmitter includes a semantic encoder, a fusion sorting encoder and a channel encoder, and receiver includes a channel decoder, a separated anti-sorting decoder, a semantic decoder.
  • Figure 3: The process of similarity semantic feature fusion.
  • Figure 4: Image restoration comparison using different sorting schemes under different bandwidth compression ratios. By sorting semantic features by sensitivity and then performing bandwidth pruning, the system can achieve adaptive semantic compression under different bandwidth constraints, effectively preserving key structural information and significantly improving image reconstruction quality.
  • Figure 5: The comparison of the reconstruction results of the same satellite image region with and without normalization. The reconstruction results obtained by normalization exhibit significantly better visual quality.
  • ...and 4 more figures

Theorems & Definitions (1)

  • Lemma 1