Deep learning for military audio classification: enhancing situational awareness with troop, artillery, jet, and UAV sounds

Read full text Summarise with VIRAT
Authors: Αλεξάνδρα Κ. Γιάννη
Cited by: 0
References: 41

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Audio-based classification of military activities enables passive monitoring and early threat detection, but remains challenging due to environmental noise and limited publicly available datasets.This thesis extends the Military Audio Dataset (MAD) by integrating 2,074 UAV recordings, creating an 8-class benchmark of 9,540 clips-the first unified dataset combining UAV detection with military audio classification.Twenty-four deep learning configurations spanning ResNets, EfficientNets, PANNs CNNs, and Audio Spectrogram Transformers were evaluated on the extended dataset.The bestperforming models, CNN14 and AST-PatchMix, achieve 94.12% accuracy, exceeding the original MAD baseline by over 3 percentage points.UAV classification proves highly reliable, with F1-scores of 0.97-0.98 and minimal confusion with helicopters or fighter jets.Transfer learning from AudioSet consistently improves performance, with gains up to 9 percentage points for transformer architectures.CNN6-Attention, a lightweight model proposed in this thesis, achieves 92.84% accuracy with only 4.1M parameters-outperforming ResNet-101 with 44.7M parameters, while remaining suitable for edge deployment in resource-constrained sensing applications.

Fields of Study & Subjects

Music and Audio Processing Speech Recognition and Synthesis Advanced Neural Network Applications

Publication details

Journal / source
University of Thessaly Institutional Repository (University of Thessaly)
Publication date
1 January 2026
Language
EN
Document type
Thesis