Comparing Optical Transfer Learning and Autoencoder Pretraining for Radar Bird-Drone Discrimination

Drone surveillance radars are dependent on reliable classification of targets for useful operation. Machine learning-based approaches, such as those based on deep learning, have been helpful for advancing such sensors' operation in real unpredictable environments. However, with labelled data being scarce resource in radar domain, machine learning-based approaches inherently carry several performance uncertainties. This is true even with approaches that involve domain translation or transfer learning, for example, with models evolved from optically trained models. In this paper, a comparison of the classification effectiveness of two different types of machine learning based approaches is performaned, namely, a convolutional neural network (CNN) pretrained with optical data, and an autoencoder-based model trained with real-world radar-only data. This comparison unlike others in the literature compares supervised and unsupervised pretraining techniques. Our results show that the unsupervised approach can outperform the resource-demanding supervised approach based on transfer learning. Furthermore, autoencoder pretraining repeated with synthetic micro-Doppler data yielded near identical classification results, which paves the possibility to utilize greater amounts of synthetic data for pretraining deep learning models. A brief inspection of the latent distribution of the simple symmetric, unregularized autoencoder confirms minor preservation of features in the learned representation.