Comparative Study of Deep Learning Model Architectures for Drone Detection and Classification

Rising popularity of commercial drone applications boosts the possibility of an enormous increase in the drone traffic within lower air space, which eventually will increase the safety threats. To deal with such potential security risks from non-cooperative drones, a precise detection of drones followed by appropriate classification is of utmost importance. Within this work we assess the performance of four different deep learning model architectures: convolutional neural network (CNN), residual neural network (ResNet), optimized radio classification through convolutional neural network (ORACLE), and long short-term memory (LSTM)-256 in terms of drone detection and classification accuracy. Raw, windowed in-phase and quadrature samples are used as input for the models. A description of the used dataset, the data processing, and the used machine learning models are provided. All four models show almost the same accuracy of more than 99% for drone detection. However, notable performance differences are observed for drone classification. The average classification accuracy of seven drone types and two noise classes using 5-fold cross-validation is relatively low for ResNet (69.92%) and CNN (72.81%) compared to ORACLE (78.74%) and LSTM-256 (80.66%).