Convolutional Neural Network Classifier for Unmanned Aerial Vehicles Detection and Identification Using Mel‐Frequency Spectrograms

ABSTRACT In the present paper, the approach to addressing the gap in the detection and identification of UAVs is presented based on a novel method in the context of Mel spectrograms for audio waves. Given the growing concern regarding the security and safety implications of drone applications, the need for focused research on the aforementioned gap is underscored. The novel and chosen solution in this context is addressed based on an employed deep neural network approach in which the establishment of a convolutional neural network is proposed, and the result is illustrated in a simulation environment to guarantee the performance of identification and classification purposes of the overall complex. Subsequently, Mel‐frequency and short‐time Fourier transform techniques are applied in order to transform the audio waveform from a specific time domain into a spectrogram in the time‐frequency domain, in a sophisticated compatibility with the real world's necessity of identification and classification of UAVs. Consequently, a 2D conventional neural network has been trained in order to achieve the learning criteria for the acoustic features from the spectrogram. Our model demonstrated significant improvements over benchmark performance, achieving a test accuracy of 98.23% for drone detection and 95.39% for drone identification, making it a promising solution for addressing UAV‐related security challenges.