Total: 1
Micro-Doppler signatures are a proven modality for discriminating between drones and birds, but their reliability degrades in low-SNR, data-constrained settings where deep learning models often fail. This paper presents a systematic study of ten statistical and physics-motivated handcrafted features for micro-Doppler classification under controlled signal degradation, using a publicly available 77 GHz FMCW radar dataset. Spectrograms are corrupted with additive white Gaussian noise, phase noise, and their combination across SNRs from -10 dB to 10 dB and phase noise levels from 1 to 10 degrees. Features are evaluated using stratified 5-fold cross-validation with Support Vector Machine and Random Forest classifiers, using fixed hyperparameters across all noise conditions. On clean data, both models achieve mean accuracy of 0.916, with F1 scores of 0.909 (SVM) and 0.892 (Random Forest). Under severe noise, entropy-based and side-lobe features remain robust, yielding F1 scores up to 0.773 and 0.831, respectively. Permutation-based importance analysis shows that some features retain complementary discriminative power even when their individual importance is low. These results highlight the value of principled feature design and provide insight into feature robustness for interpretable radar classification systems.