Improving Micro-Doppler Target Classification by Fusing Spectrograms and Additional Target Features

Deep learning has revolutionized radar target classification. A common approach is a convolutional neural network (CNN) model and an image representation of the target, such as a micro-Doppler spectrogram. By using only a spectrogram to represent a target, however, valuable information about the target is omitted. In this paper, we demonstrate a fused CNN-and feature-model that combines spectrograms with additional target features, namely radar cross section (RCS), range-change and azimuth-change. The efficacy of the approach is shown by comparing the combined model with feature-only, spectrogram-only, and ensemble models on two different datasets. The combined model shows performance improvements, particularly in the lower-scoring classes. More importantly, we also explore how the combined model utilizes and balances the additional features and spectrogram by investigating the attribution score of each part of the model. The attribution analysis reveals that the models have learned to depend on each other and that the balance is highly dependent on class and signal to noise ratio (SNR).