Wavelet packet convolutional network for low-altitude target trajectory classification

The rapid growth of the low-altitude economy and the proliferation of unauthorized UAVs demand accurate identification of “low-slow-small” targets, including multirotor drones, fixed-wing UAVs, balloons, and birds. This paper extends the Wavelet Packet Convolutional Network (WPCN)—originally developed for infrared dim-target recognition—to trajectory-based low-altitude object classification. Multi-dimensional trajectory features (position, velocity, heading, acceleration—8 dimensions \(×\) 256 frames) are decomposed via per-channel Wavelet Packet Transform (WPT) into a \(75 × 32\) frequency-domain representation, enabling a residual 1D CNN to extract multi-scale motion pattern features. Depthwise separable convolution and batch normalization removal strategies are introduced for edge deployment in anti-UAV systems. Experiments on two benchmarks—a self-constructed 5-class low-altitude trajectory dataset (LOA) and 7 UCR public time series datasets—demonstrate that WPCN achieves \(99.63% ± 0.19%\) mean accuracy (10 runs) on LOA under 1 m measurement noise, significantly outperforming LSTM ( \(99.33%\) ), Transformer ( \(99.00%\) ), and ROCKET ( \(99.00%\) ). The lightweight WPCNLG variant (54 K parameters, 61 s training) approaches LSTM-level accuracy at a fraction of the cost. WPT ablation shows \(+0.30\) pp accuracy gain with \(2.4×\) training speedup. DBA and DTW analyses confirm a \(19.6%\) improvement in the inter-class separation ratio after WPT transformation. Friedman tests confirm statistically significant differences across models on the UCR benchmark ( \(p = 0.010\) ).