Interference Analysis and Detection in Space-Air-Ground Integrated Networks

Space-air-ground integrated networks (SAGIN) encounter highly heterogeneous interference arising from satellite, aerial platforms, and terrestrial links, while the scarcity and imperfection of labeled data required for spectrum sensing and interference detection further complicate the associated analysis. To address these challenges comprehensively, this paper presents two complementary contributions. First, the effects of the interference, caused by the intra-satellite interference from direct-to-satellite terrestrial users (STUs), inter-satellite interference of direct-to-satellite unmanned aerial vehicles (SUAVs) and of STUs, as well as from terrestrial users (TUs) accessing terrestrial base stations, are specially analyzed for the typical urban, suburban, and urban-suburban fringe scenarios. Moreover, the carrier-to-interference plus noise ratio are also investigated in the presence of the four-color frequency reuse, the full-frequency reuse and the dynamic-frequency reuse schemes. Second, a real-time semi-supervised interference detection method is conceived. This method employs the confidence-based label cleaning during the pretraining phase and the uncertainty-aware pseudo-label selection during the training phase in the teach-student learning (CleanUPS-Teacher) framework. The conceived CleanUPS-Teacher can also mitigate effectively the negative impacts of labeling errors and pseudo label noise on interference detection. Experimental validations on real satellite spectrum datasets demonstrate that CleanUPS-Teacher achieves comparable or superior performance relative to fully supervised approaches using only 40% labeled data. Moreover, the proposed method maintains robust performance, outperforming existing semi-supervised benchmarks in terms of accuracy and stability, even under the condition of labeling noise levels as high as 15%.