An FMCW radar target recognition method based on multi-strategy improved PO-SVMD

To address clutter interference with micro-motion features and parameter selection difficulties in L-band radar echoes from unmanned aerial vehicles (UAVs), this study proposes a recognition method combining a multi-strategy parrot optimizer (MS-PO), successive variational mode decomposition (SVMD), and a bidirectional gated recurrent unit (BiGRU). First, piecewise chaotic initialization, adaptive crossover, and stagnation perturbation improve the parrot optimizer, which determines the SVMD penalty factor by minimizing envelope entropy. SVMD then extracts echo modes successively. Micro-Doppler bandwidth, spectral centroid, and log-transformed true energy are extracted window by window and arranged chronologically into a 32 × 3 matrix. The BiGRU learns temporal changes in these features to identify four UAV classes. On L-band echoes from LSS-FMCWR-2.0, mean accuracy reaches 94.53% over three repetitions of stratified ten-fold cross-validation, exceeding SVMD-BiGRU with a fixed penalty factor by 8.53 percentage points. Track-level leave-one-out evaluation using recording groups achieves 88.10% accuracy. With the same SVMD decomposition and BiGRU classifier, MS-PO improves accuracy over the standard parrot optimizer by 0.78 and 8.83 percentage points under the two protocols, respectively.