Distributed Multimodal Split Inference Over LEO Satellite Networks via VQ-Based Semantic Transmission

Abstract

Distributed multimodal inference over low Earth orbit (LEO) satellite networks can enhance the timeliness of Earth-observation services, yet bandwidth-limited inter-satellite links make dense feature exchange prohibitively costly. This paper proposes a distributed multimodal split-inference framework with vector-quantized (VQ) semantic transmission, where sensing satellites process local modalities with lightweight front-end subnetworks and transmit only discrete semantic indices to a computing satellite for feature reconstruction, multimodal fusion, and final inference. A steady-state pipeline latency model is further developed to support resource-aware split-point and codebook selection. Experiments on RGB-SAR semantic segmentation using the DDHR dataset show that the optimal split varies with sensing-side capability and ISL bandwidth. Under the default setting, the proposed method achieves 95.6% mIoU with a $\mathbf{9 ~ K B}$ payload and $\mathbf{2 1. 0 8 ~ m s}$ steady-state latency, delivering a 9.95× speedup over raw-image transmission while maintaining competitive accuracy. These results demonstrate the effectiveness of VQ-based semantic transmission for communication-efficient in-orbit collaborative inference.