UAV passive sensing method based on hierarchical cross‑attention network

To address weak echo submersion and multipath interference in digital terrestrial multimedia broadcast (DTMB) external illuminator, a UAV passive sensing method based on hierarchical cross-attention network (HCANet) was proposed. A dual-branch architecture was constructed to extract temporal dependencies via time-domain multi-scale convolution and micro-Doppler features via frequency-domain amplitude-phase joint encoding. Hierarchical cross-attention modules were utilized to adaptively align and fuse time-frequency features in the feature space. Experimental results show that the method maintaines robustness within -10 to 10 dB signal-to-noise ratios. The detection accuracy under 10 dB additive white Gaussian noise, Rician, and Rayleigh channels reaches 95.75%, 99.50%, and 71.25%, respectively, and the accuracy of measured data in complex urban environments reaches 99.65%. These results validate the effectiveness in low signal-to-noise ratio and strong-scattering environments.