Bird and micro-drone spectral width and classification

This paper reports on the class separability of spectrograms featuring bird and micro-drone targets produced by an L-Band staring radar. Multi-rotor drones with small propellor blades are less likely to show strong micro-Doppler sidebands depending on the range and operating frequency. With this, we were incentivized to measure the separability of the target classes relying only on the body Doppler information captured in the spectrograms. A spectral width feature extraction method was tested using both a set of single drone and bird targets, as well as a larger dataset including spectrograms containing multiple targets and a mixture of classes. These features were employed to inform a simple classifier yielding an 83% classification accuracy in the single target case. The results were then compared to a convolutional neural network baseline that achieved 89% accuracy on the larger, more complex dataset.