Hybrid Semantic-Bit Communications for NTN Downlinks: A Hierarchical Multi-Agent DRL Approach

Sixth-generation (6G) networks target intelligent, context-aware communications, with Non-Terrestrial Networks (NTNs) essential for ubiquitous global coverage. However, limited bandwidth, severe path loss, massive data traffic, and dynamic, harsh channel conditions challenge both reliability and efficiency. This paper proposes a two-hop NTN downlink framework that synergistically integrates Semantic Communication (SemCom) and Bit-based Communication (BitCom) for efficient image delivery from a Low Earth Orbit (LEO) satellite to ground users via High-Altitude Platforms (HAPs). A deep learning-enabled architecture is developed, featuring a semantic-aware encoder at the LEO satellite and a hybrid semantic–bit processing unit at the HAP. Exploiting full channel state information (CSI), semantic relevance, and computational capability, the HAP dynamically categorizes users into Semantic-Aware Users (SAUs) and Bit-Based Users (BBUs), forwarding semantic features to SAUs or reconstructing bit-level data for BBUs. To holistically evaluate performance, we introduce a novel Universal Performance Metric for Hybrid Semantic and Bit-based Communication (UPM-HSBC), termed Hybrid Communication Effectiveness (HCE), which jointly captures semantic relevance, spectral efficiency, and channel dynamics. Building on this, we formulate a long-term HCE maximization problem and design a Hierarchical Multi-Agent Twin Delayed Deep Deterministic Policy Gradient for Semantic–Bit Resource Allocation (HMATD3-SBRA) algorithm. The framework jointly optimizes transmit power, bandwidth, computational resources, and mode selection across LEO satellite–HAP–ground links, while explicitly capturing inter-agent dependencies. Extensive simulations reveal that our approach achieves up to 18% higher spectral efficiency and 22% reduction in latency, while effectively reducing mode-switching overhead by 15%. It further sustains 10–25% HCE gains across diverse signal-to-noise ratio (SNR) regimes and resource constraints, validating the framework’s effectiveness for semantic-aware NTN downlinks in 6G.