Abstract
With the rapid proliferation of unmanned aerial vehicles (UAVs), reliable identification based on radio frequency (RF) signals has become increasingly important for both civilian and security applications. This paper proposes a spatiotemporal feature extraction and classification framework based on bispectral analysis. Specifically, bispectral estimation is used to convert one-dimensional RF signals into two-dimensional bispectrum feature maps that capture higher-order spectral characteristics and nonlinear dependencies. Based on these characteristics, a two-stage network was constructed for spatiotemporal feature extraction and classification. The first stage utilizes a ResNet18 network to extract spatial structural features from individual bispectrum maps. The second stage employs an LSTM network to learn temporal dependencies across the sequence of bispectrum maps, capturing the continuity and evolution of signal characteristics over time. The experimental results on a public dataset of UAV RF signals show that this method improves recognition accuracy by 6.78% to 13.89% compared to other existing methods across five categories of UAVs.