Augmentation Based on Spectrogram Segments for UAV Operating Channel-Robust CNN Classifiers

The convolutional neural network (CNN) is effective to classify radio frequency (RF) signals of unmanned aerial vehicles (UAVs), although the variation of UAV operating channels can degrade the performance of the CNN. As the CNN is data hungry, an intuitive solution is to capture UAV signals of all channels to train the CNN. However, the signal collection is time-consuming and expensive. Hence, this paper proposes a data augmentation scheme based on the frequency characteristics of UAV signals, which approximately simulates UAVs operating on different channels. With this scheme in the training pipeline, we use signals of a single UAV channel to train CNNs that can classify UAVs operating on arbitrary channels. Extensive indoor and outdoor experiments are conducted, and the collected signals are also released as part of the technical contributions of our work. The experimental results show that the proposed data augmentation scheme can improve the classification accuracy of CNNs by 60%.