A comparative evaluation of spectrogramand bispectrumbased transfer learning for drone audio classification
Abstract
Acoustic signal analysis serves as an essential approach for detecting drones operated via optical fibers, which evade conventional radio-frequency detection methods. Nonetheless, achieving reliable acoustic-based detection poses notable challenges due to the pronounced sensitivity of classification models to environmental noise and low signal-to-noise ratios. While deep transfer learning has emerged as a powerful tool for audio classification, the choice of optimal time-frequency representation remains a critical factor in model performance. This paper presents a comparative evaluation of spectrogramand bispectrum-based transfer learning for drone audio classification. Utilizing a comprehensive dataset of drone and non-drone acoustic signals, raw audio data is transformed into standard 2D spectrograms and higher-order bispectrum representations, which are then classified using three pre-trained models: ResNet-18, ResNet-50 and ViT Base 16. The results of this analysis suggest that higher-order spectral analysis provides a more robust and dependable feature representation in case of smaller model architectures, while larger networks struggle to generalize on the nonlinear features of the bispectrum and favor spectrogram representations.