Hierarchy-Aware and Channel-Adaptive Semantic Communication for Bandwidth-Limited Data Fusion

Lei Guo, Wei Chen, Yuxuan Sun, Bo Ai, Nikolaos Pappas, Tony Q. S. Quek

Abstract

This paper introduces a hierarchy-aware and channel-adaptive semantic communication framework designed to reconstruct high-resolution hyperspectral images (HR-HSI) while significantly reducing bandwidth requirements. By intelligently fusing deep structural features with shallow spatial details and utilizing a Transformer-based attention mechanism to prioritize data transmission, the proposed framework achieves a 66% reduction in bandwidth with a marginal loss in reconstruction quality (less than 1 dB). The system demonstrates robustness across various channel conditions and outperforms traditional compression-based methods.

Introduction

Hyperspectral imaging (HSI) is a critical tool in remote sensing and environmental monitoring, providing rich spectral data across numerous bands. However, a persistent hardware trade-off exists: sensors typically capture high spectral resolution at the expense of spatial resolution, resulting in low-resolution hyperspectral images (LR-HSI). To overcome this, researchers employ data fusion, combining LR-HSI with high-resolution RGB (HR-RGB) images to synthesize high-resolution hyperspectral data.

The primary bottleneck for deploying these systems on drones or satellites is bandwidth. Traditional fusion techniques generate massive data volumes that often exceed the size of the original inputs, making real-time transmission impractical. Conventional compression methods, such as BPG, often fail because they do not account for the semantic importance of specific features relative to the final reconstruction task. This research addresses these challenges by proposing a Semantic Communication (SC) framework that prioritizes the transmission of "meaningful" features over raw pixel data.

Methodology

The proposed architecture shifts the communication paradigm from bit-level transmission to feature-level semantic exchange. The framework is built upon three core innovations:

1. Hierarchical Feature Extraction

The system employs specialized modules to extract spectral features from the LR-HSI and spatial features from the HR-RGB. Crucially, it distinguishes between "deep" features, which represent the overall global structure, and "shallow" features, which capture fine-grained spatial details. This separation allows the system to manage different levels of information density according to their importance for reconstruction.

2. Hierarchy-Aware Fusion Module

Acting as the core of the transmitter, this module utilizes a Transformer-based attention mechanism. By employing a sequence of Self-Attention (SA) and Cross-Attention (CA), the module generates dynamic guidance vectors. These vectors act as "masks" that determine the optimal combination of deep and shallow features, creating a compact, high-utility semantic stream ($S$).

3. Channel-Adaptive Transmission and Decoding

The system adaptively integrates features based on current channel conditions. Instead of transmitting the entire feature set, it sends the compressed semantic representation alongside small mask vectors. These masks serve as a "recipe" for the receiver, occupying negligible bandwidth. At the receiver, a trainable fusion decoder uses these features and masks to "unfold" the data back into a high-resolution hyperspectral image.

Experimental Evaluation

Datasets and Setup

The model was benchmarked using two standard hyperspectral datasets:

  • CAVE Dataset: 31 spectral bands with a 512 × 512 spatial resolution.
  • Washington DC Mall Dataset: 191 spectral bands with a 128 × 128 spatial resolution.

Performance Metrics

The framework demonstrated significant improvements over baseline methods across several dimensions:

  • Bandwidth Efficiency: The method achieved a 66% (two-thirds) reduction in bandwidth consumption compared to "Full Fusion" methods, while maintaining a reconstruction quality drop of less than 1 dB.
  • Reconstruction Accuracy: On the CAVE and DC Mall benchmarks, the framework achieved up to a 2 dB improvement in Peak Signal-to-Noise Ratio (PSNR) compared to single-source transmission strategies.
  • Robustness: The system remained effective across diverse channel models, including AWGN, Rayleigh, and MIMO channels, consistently outperforming conventional "BPG+Capacity" schemes.

Discussion and Limitations

While the framework offers a significant leap in efficiency, it faces challenges in extremely noisy environments. In low Signal-to-Noise Ratio (SNR) scenarios, the channel-adaptive mechanism can degrade as the quality of received features becomes too low for the attention mechanism to process effectively. The authors suggest that future iterations could integrate diffusion-based models to "hallucinate" or recover missing details in high-noise conditions.

Additionally, the computational requirements are noteworthy. The model requires approximately 486.49 GFLOPs with a latency of 0.11 seconds. While this allows for near-real-time performance, it remains a significant load for power-constrained edge devices like small UAVs. Future research will likely focus on model distillation to reduce the computational footprint for deployment on lower-power hardware.

Conclusion

By moving beyond traditional bit-level compression and focusing on hierarchy-aware semantic features, this framework provides a viable path for high-quality hyperspectral data fusion in bandwidth-constrained environments. The use of Transformer-based masks allows for a drastic reduction in data transmission without sacrificing the structural integrity of the reconstructed images.

Summary updated

BibTeX
@article{guo2025hierarchy,
  author = {Lei Guo and Wei Chen and Yuxuan Sun and Bo Ai and Nikolaos Pappas and Tony Q. S. Quek},
  title = {{Hierarchy-Aware and Channel-Adaptive Semantic Communication for Bandwidth-Limited Data Fusion}},
  journal = {IEEE Wireless Communications Letters},
  year = {2025},
  doi = {10.1109/lwc.2025.3554572},
}
Figure