Effective Bandwidth Analysis for Acoustic UAV Detection Using Log-Mel Spectrograms

Abstract

Reliable acoustic UAV detection requires preserving discriminative rotor and motor signatures while limiting non-informative spectral content. However, the effective bandwidth required by deep learning-based detectors remains insufficiently understood independently of sampling rate, feature extraction, and model architecture. This study evaluates acoustic bandwidth using log-Mel spectrograms extracted from native 96 kHz multi-microphone recordings. In the main controlled ablation, only the upper filterbank limit was varied from 2 to 48 kHz, while the STFT parameters, feature dimensions, model architecture, dataset splits, and training procedure remained fixed. Supplementary 1–8 and 2–8 kHz configurations examined the effect of excluding the lowest-frequency content. Bandwidth affected detection performance non-monotonically. The 0–8 kHz configuration achieved the highest observed F1-score of 0.921 and precision of 0.908, but its F1-score was not statistically different from that of 0–10 kHz $(p=0.645)$. The full 0–48 kHz configuration achieved the highest recall of 0.953 but lower precision. Thus, 0–8 and 0–10 kHz provided comparable balanced performance under the evaluated conditions.