Radar sensing offers a method of achieving 24-h all-weather drone surveillance, but in order to be maximally effective, systems need to be able to discriminate between birds and drones. This work examines drone-bird classification performance as a function of signal to noise ratio (SNR). Classification at low SNR values is necessary in order to classify drones with a small radar cross-section (RCS), as well as to facilitate reliable classification at longer ranges. To investigate the relationship between classification performance and SNR, Gaussian noise is added to an experimentally obtained dataset of radar spectrograms. Classification is performed by convolutional neural networks (CNNs). It is shown that for the data available classification accuracy drops with falling SNR, as might be expected for any given CNN. The degree to which performance degrades with reduced SNR is presented. It is further shown that simpler network architectures are more robust to noise. Finally, it is demonstrated that data augmentation can be used as a means of enhancing classification accuracy at lower SNR values. Bayesian optimisation is used to find the optimal augmentation hyperparameters and overall, classification accuracies of 92% are achieved at low SNR.This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited.
It has already been shown that multibeam staring radar is able to detect and track low observable targets such as drones due to its high sensitivity [1]. Due to this level of sensitivity, targets that have a similar RCS to drones are also detected and tracked. These are predominantly birds. Birds and drones are similar in several ways such as flight altitude, velocity and manoeuvrability [2] such that discrimination between them is challenging. Hence, there is a need to look for high performing methods of classification, for example, machine learning. Supervised training of machine learning classifiers requires accurately labelled training data. For control targets, such as drones, truth data from the onboard GPS logging can be used for data labelling. However, opportune bird targets require a separate data collection method that enables association with the radar output for a classifier to be effectively trained. This paper shows a method of collecting and displaying ground-truth for small targets onto GoogleEarth so that the radar data can be appropriately used to create accurate training data for a machine learning, drone and bird classifier. Results of classification performance are presented showing high performance that is aided by the availability of more effective truth data.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.