Robust Radar Micro-Doppler Target Classification of Small Drones by Data Augmentation

The presence of small drones, which are desirable to detect and classify for security reasons, is an increasing problem. Classification of radar targets frequently concerns scenarios with high Signal to Noise Ratios (SNRs) and large amounts of data. However, such optimal conditions are not guaranteed, which actualizes robustness studies. This paper uses (real) radar measurements to study the robustness of the cases to reduce the SNRs and increase the data scarcity for the problem of micro-Doppler target classification of birds and small drones with Convolutional Neural Networks (CNN). Investigations for improving the classification robustness included three data augmentation methods to enhance the amount of training data by adding synthetic data. The investigated methods contribute to the development of a more robust CNN classifier by improving validation classification results and retaining performance while reducing the SNRs and the amount of training data. Notably, promising results were obtained by stochastically combining the methods of window warping and window slicing.