PaperlayerView exact revision on arXiv

Source-linked AI summary

Acoustic UAV Detection in Battlefield Scenarios: Handling Noise, Domain Shift, and Weak Labels

Vadym Vilhurin, Volodymyr Sydorskyi, Andrii Shevtsov

arXiv:2608.14287v1cs.SDcs.AIcs.CV

TL;DR

Acoustic UAV detection remains difficult under realistic battlefield noise, heterogeneous recording devices, and cross-domain shift. This paper combines PCEN, attention-based pooling, and domain-aware training with multi-microphone data and reports consistent improvements in cross-device generalization and detection accuracy.

  • Problem

    Acoustic UAV detection remains underexplored across heterogeneous recording devices and is limited by environmental interference, hardware variation, and cross-microphone generalization challenges.

  • Method

    The framework combines PCEN spectrograms, attention-based pooling, and domain-aware training with multi-microphone data and noise-driven curriculum augmentations.

  • Results

    The approach shows consistent improvements in cross-device generalization and detection accuracy across multiple open-source and proprietary baselines.

  • Takeaways & Limitations

    Including auxiliary UAV classes from an auxiliary domain improves generalization to the target UAV class in the target domain.

Abstract

from arXiv · show

Passive acoustic sensing offers a critical, cost-efficient, and, crucially, passive alternative for detecting small unmanned aerial vehicles. However, the practical deployment of acoustic systems is discouraged by extreme environmental noise and sensor-induced domain shift caused by heterogeneous hardware. This paper addresses these challenges by introducing a robust framework optimized for real-world battlefield conditions. We propose the integration of Per-Channel Energy Normalization (PCEN) and attention-based pooling to enhance feature extraction under low signal-to-noise ratio scenarios. We further propose a domain-aware training strategy that leverages auxiliary classes and multi-microphone data to mitigate cross-domain performance degradation. Evaluated on a unique dataset of combat-zone recordings from the Ukrainian frontlines, our approach significantly outperforms existing baselines, increasing the F1 score from 55.4% to 78.6%. This paper was originally presented at the International Conference on Military Communication and Information Systems (ICMCIS), organized by the Information Systems Technology (IST) Scientific and Technical Committee, IST-224-RSY - the ICMCIS, held in Bath, United Kingdom, 12-13 May 2026.

I. INTRODUCTION · II. RATIONALE TO MILITARY APPLICATION

The paper frames acoustic UAV detection as a passive, low-cost approach for contested military environments, while addressing severe environmental noise, weak labels, and sensor-induced domain shift. It proposes PCEN, attention-based pooling, and domain-aware training using multi-microphone and auxiliary-class data from Ukrainian frontlines.

  • I. INTRODUCTION: Small UAV proliferation threatens critical infrastructure, personnel safety, and operational secrecy, making early and robust detection essential in realistic battlefield conditions.These conditions include strong environmental interference, clutter, and heterogeneous sensing hardware.
  • I. INTRODUCTION: Existing UAV detection methods use vision, radio-frequency sensing, radar, and infrared, but face operational limitations including occlusion, interference, protocol dependence, and costly hardware.The limitations are especially relevant outside controlled settings.
  • I. INTRODUCTION: Acoustic detection is passive, low-cost, protocol-agnostic, and largely invariant to visual occlusion, making it attractive for early detection in cluttered and contested environments.Its passive nature also reduces susceptibility to radio jamming or electronic countermeasures.
  • I. INTRODUCTION: Real-world acoustic detection is challenged by variable noise from wind, vehicles, machinery, and human activity, which can obscure UAV signatures and degrade performance.Data from different microphones and deployment locations also exhibit substantial sensor-induced domain shift.
  • I. INTRODUCTION: The paper investigates acoustic UAV detection under realistic noise, strong sensor-induced domain shift, and data scarcity as a combined practical setting.The challenge spans recording-device configuration, sensor location, background noise, and target-class differences across two acoustic domains.
  • I. INTRODUCTION: The proposed CNN framework combines PCEN spectrograms with attention-based pooling to improve robustness to environmental noise and weakly labeled data.PCEN denotes Per-Channel Energy Normalized spectrograms.
  • I. INTRODUCTION: Domain-aware training uses multiple microphone domains with noise-driven and curriculum-based augmentations to mitigate cross-domain shift.Including auxiliary UAV classes from the auxiliary domain improves generalization to the target UAV class in the target domain.
  • II. RATIONALE TO MILITARY APPLICATION: Domain-shift mitigation is operationally important because changing seasons, new UAV acoustic signatures, and sound mimicry can cause missed detections or false alarms.The paper emphasizes this issue in modern warfare, including the Russia-Ukraine war and NATO border surveillance.

III. RELATED WORK

Prior work establishes acoustic UAV detection with classical spectral methods and increasingly robust CNN- and multimodal approaches, but real-world deployment challenges remain insufficiently addressed. Key gaps include environmental interference, cross-microphone domain shift, extreme imbalance, and scarce operational data.

  • Acoustic UAV detection uses mechanical signatures, while early spectral-feature and Support Vector Machine systems separated drones from background noise under controlled conditions.
  • CNN-based systems improved robustness over handcrafted features and enabled noisy outdoor recognition, although strong wind and mechanical noise still degrade performance.
  • Multimodal approaches such as TRIDENT combine audio with video, radio-frequency, or infrared video modalities to achieve more robust drone detection.
  • Existing studies leave cross-microphone generalization, extreme class imbalance, and domain data scarcity largely unaddressed despite their importance in real-world UAV monitoring.
  • CNNs, EfficientNet, ConvNeXt, and CRNNs provide established architectures for spectrogram classification and weakly labeled sound event detection.
  • PCEN provides adaptive gain control and stationary-noise suppression, but it has not been systematically evaluated for UAV detection under extreme domain imbalance.

IV. DATASET

The dataset contains more than 300,000 nine-second recordings from two microphone domains, with device-aware splits designed to prevent leakage. Its classes and label distributions differ substantially across devices, including a rare Small drone class in the small set.

  • Dataset construction: More than 300,000 recordings are nine seconds long and sampled at 32 kHz, with clips from each continuous stream kept within a single split.Adjacent clips may contain overlapping or highly similar acoustic content, so stream-level splitting prevents data leakage.
  • Acoustic domains: The two microphone types capture different conditions and hardware responses, including tower-mounted Mic-1 recordings with vibration, wind, animal, and frequency-response effects.Mic-2 is described as the second distinct microphone type and acoustic domain in the supplied passage.
  • Acoustic domains: Mic-2 is the target acoustic domain, while Mic-1 is an auxiliary domain; validation and test sets include both devices and their corresponding classes.Training may use Mic-2 alone or both microphone types.
  • Label distribution: The small set has a highly imbalanced device distribution: Mic-1 provides more than 99% of training data, while Mic-2 contributes approximately 3,000 samples (< 1%).Only 441 samples belong to Small drone, which is absent from Mic-1, creating a partially disjoint label space.

V. VALIDATION PROTOCOL

The validation protocol uses fixed, time- and location-independent splits instead of K-fold cross-validation, with seasonal and geographic separation between development and final evaluation. Model selection uses Macro F1 across all classes while final reporting also includes binary detection metrics.

  • Data Splits: Models are evaluated on fixed validation and test splits based on time and location, with all three sets completely time independent.This avoids K-fold cross-validation to reduce computational cost.
  • Data Splits: The Training Set consists of controlled field-test recordings collected in non-combat zones during development.It represents the initial data available for model development.
  • Data Splits: The Validation Set contains Winter recordings from active combat zones and field tests for hyperparameter tuning and overfitting checks.Validation and test class distributions are depicted in Table II.
  • Data Splits: The Test Set contains Spring recordings exclusively from active combat zones, evaluating performance in a new season and unseen locations.It serves as the final evaluation set.
  • Model Selection and Metrics: The best checkpoint is selected by Macro F1 across all classes, while Precision, Recall, and F1-score are reported for binary Small Drone detection.Tracking the binary metric helps prevent overfitting to auxiliary classes and addresses target-class scarcity in validation.

VI. DOMAIN SHIFT

The domain-shift analysis finds clear microphone-related separation, while seasonal shift appears absent and environmental noise remains distinct from other classes. Enhanced PCEN reduces noise-class separability, but quantitative results indicate that sensor-domain alignment is not fully resolved.

  • VI. DOMAIN SHIFT: t-SNE embeddings show Mic-2 environmental noise forming a distinct cluster, demonstrating domain shift between microphone domains.Noise from the two domains is partially separated but remains well separated from other classes.
  • VI. DOMAIN SHIFT: Enhanced models struggle to align sensor domains, with clusters appearing more distinct than for the baseline model.This conclusion comes from Silhouette Coefficient and k-NN cluster-purity evaluations in the Global Noise Sensor Shift experiment.
  • VI. DOMAIN SHIFT: Seasonal-shift comparisons suggest no seasonal shift for the signals.The analysis compares validation and test distributions.
  • VI. DOMAIN SHIFT: Sil. 0.2592 and k-NN purity 93.91% for Enhanced (PCEN) reduce noise-class separability compared with the baseline’s Sil. 0.3851 and k-NN > 98%.PCEN blurs non-stationary environmental textures, suggesting greater cluster overlap.
  • VI. DOMAIN SHIFT: The baseline achieves ROC AUC close to 100% on both validation and test sets in adversarial noise-source classification.This indicates near-perfect separability of samples recorded with different sources.

VII. METHOD

The method adapts a ConvNeXt-Tiny and SED-style attention architecture for weakly labeled UAV detection, replacing destructive frequency pooling and modifying attention outputs for sharper, mutually exclusive multiclass predictions.

  • Architecture: The framework uses a ConvNeXt-Tiny backbone with an SED-style attention head to accommodate weak labels.This architecture follows the BirdCLEF+ 2025 second-place solution.
  • Low-SNR feature extraction: A learnable convolutional frequency projector replaces GeM pooling to preserve discriminative spectral resolution under low-SNR noise.The change addresses masking caused when high-energy low-frequency noise overlaps the UAV signal.
  • Multiclass classification: Removing tanh from attention permits unbounded scores that sharpen temporal focus in multiclass classification.The modification allows the attention mechanism to emphasize relevant time regions more strongly.
  • Multiclass classification: A class-wise softmax replaces the final sigmoid to enforce mutual exclusivity among classes.The softmax is applied to the aggregated output.

A. Audio Representation · B. Data Augmentation

The framework uses PCEN spectrograms to improve robustness to environmental noise and applies targeted waveform and spectrogram augmentations to address imbalance and cross-microphone domain differences. These augmentations preserve class information while increasing variability and forcing attention beyond the loudest signal segment.

  • A. Audio Representation: All inputs are converted into PCEN spectrograms to address environmental dynamic-range variance.PCEN suppresses stationary background noise and normalizes loudness variations, improving weak spectral structures under high-energy low-frequency noise.
  • B. Data Augmentation: Broad waveform augmentations address severe data imbalance, particularly in Mic2.The augmentation suite operates directly on the waveform.
  • B. Data Augmentation: Spec augment-style masking introduces temporal and amplitude variability without drastically altering underlying spectral signatures.This increases variation while retaining the core spectral characteristics of the audio.
  • B. Data Augmentation: Noise injection via mixup mixes noise into training audio while preserving the original class label.It generates additional minority-domain examples and gradually pulls Mic-1 representations toward the Mic-2 acoustic space.
  • B. Data Augmentation: Randomly replacing 0.5–1.0 seconds inside the highest-energy 2-second region prevents reliance on the loudest signal chunk.This augmentation applies to all non-noise classes and follows a curriculum increasing replacement duration from 0.3–0.5 s to 0.5–1.0s.
  • B. Data Augmentation: The loudest-segment removal curriculum forces the model to use features beyond the maximum-amplitude region.Because the signal is almost guaranteed to exist around that region, masking encourages reliance on complementary evidence.

C. Training Setup · VIII. EXPERIMENTS AND DISCUSSION · A. External Baseline Models

The training setup combines recordings from both microphones, using Mic-1 classes as auxiliary supervision while optimizing primary Small drone versus all-other-class discrimination. Experiments compare several open-source or open-weight baselines, including zero-shot segmentation, a large UAV-specialized transformer, and a lightweight audio detector.

  • C. Training Setup: The model trains on the combined dataset from both microphones, with Mic-1 classes serving as auxiliary classes.This auxiliary setup supports learning across heterogeneous microphone data.
  • C. Training Setup: The backbone and attention head learn generalizable acoustic structure while optimizing Small drone versus all other classes.The primary task remains binary discrimination between Small drone and all other classes.
  • A. External Baseline Models: Several open-source and open-weight models are used as-is as experimental baselines.These models provide external reference points for the experiments.
  • A. External Baseline Models: SAM Audio assesses zero-shot UAV detection using the prompt Small UAV FPV flight and FP16 inference.Detected target segments are converted into clip-level decisions by thresholding maximum absolute amplitude.
  • A. External Baseline Models: AST-Drone is a large AST variant fine-tuned specifically for UAV detection.It performs best among the open-source baselines but lacks real-world training data and cross-environment generalization.
  • A. External Baseline Models: TRIDENT LeNet-Audio evaluates standalone performance of a compact, resource-efficient UAV detector against larger specialized approaches.Its compact size prevents it from handling drone detection scenarios effectively.

B. Detailed Training Setup · C. Experiments

The training setup addresses extreme imbalance, harsh acoustic variability, and heterogeneous domains through resampling, augmentation, specialized preprocessing, and domain-aware inputs. Experiments show that auxiliary classes, noise-focused regularization, frequency-aware processing, and complementary PCEN/Mel modalities improve robustness, while target-only training collapses.

  • B. Detailed Training Setup: Target-class prevalence below 1% risked trivial noise-detector collapse, so training used square-root sampling with replacement and independently sampled environmental-noise domains.Both domains shared one noise label but were sampled according to their respective counts.
  • B. Detailed Training Setup: Augmentations simulated harsh target conditions through ±100ms time shifts, sinusoidal warping, ±3dB gain changes, waveform reversal, and domain-specific noise injection.Noise mixing used source or target noise with equal probability, capped λ at 0.3, and retained the original label to synthesize low-SNR examples.
  • B. Detailed Training Setup: The implementation used a pretrained ConvNeXt-Tiny backbone, 32 kHz nine-second inputs, spectrograms, and higher-resolution 4096-sample, 256-band PCEN/Log-Mel features.Log-Mel baselines instead used 2048-sample windows and 128 Mel bands.
  • C. Experiments: Target-only training on Mic-2 caused model collapse, demonstrating that the limited target-domain data was insufficient for generalization without domain adaptation.
  • C. Experiments: Auxiliary aircraft classes improved discrimination by providing overlapping frequency structures and richer training representation, while potentially causing negative transfer when spectral overlap was unhelpful.The identified risks included unrelated-feature capacity use, diluted target confidence, and domain-specific-noise overfitting.
  • C. Experiments: Noise-injection Mixup aligned source and target noise distributions, while curriculum RMS masking reduced reliance on the highest-energy segment by exposing weaker harmonics across clips.
  • C. Experiments: The frequency projector preserved harmonic-ratio information under low-frequency interference by learning frequency-axis features instead of collapsing frequency information through global averaging.This addressed cases where the fundamental drone frequency was masked.
  • C. Experiments: 82.22% F1 was achieved by the full-dataset ensemble, while the small-dataset model achieved 81.14% F1 and outperformed the Zvook baseline by a large margin.PCEN suppressed stationary wind rumble, complementing Mel-spectrogram sensitivity to gain.

D. Detection Range Analysis

The final ensemble model was evaluated across human-annotated perceived-distance subgroups using ROC curves and ROC AUC. It achieved high performance for nearby or medium-distance and far-distance flybys, despite lower signal-to-noise ratios and weaker acoustic signatures at greater distances.

  • Distance Subdivision: Human labelers subdivided the Small drone class by perceived distance, providing the only practical approximation of sensor-to-drone distance in real combat scenarios.Perceived distance may be influenced by physical obstacles, microphone characteristics, and drone type.
  • Evaluation Method: The final ensemble model was evaluated across distance ranges using ROC curves and the ROC AUC metric.The analysis used the nearby, medium-distance, and far-distance flyby subgroups.
  • Performance by Range: 97.2% ROC AUC was achieved by 2304 nearby or medium-distance flyby recordings.This subgroup combined recordings labeled as nearby or medium-distance flybys.
  • Performance by Range: 93.8% ROC AUC was achieved by 241 far-distance flyby recordings.The lower performance for distant flybys was attributed to lower signal-to-noise ratios and weaker acoustic signatures at larger distances.
  • Robustness Across Ranges: The model maintained high detection accuracy across nearby or medium-distance and far-distance subgroups, indicating robust performance across varying detection ranges.Performance declined for distant flybys, but remained high across both evaluated subgroups.

IX. CONCLUSION

The proposed domain-aware framework combines PCEN representations, attention-based pooling, and noise-driven curriculum augmentations to improve cross-device generalization and detection accuracy. The conclusion identifies limitations and future directions involving finer UAV taxonomies, sequential modeling, spatial metadata, and adaptation to data-scarce frontline locations.

  • Contributions: The framework combines PCEN representations, attention-based pooling, and noise-driven curriculum augmentations for domain-aware acoustic UAV classification.Evaluations against multiple open-source and proprietary baselines showed consistent improvements in cross-device generalization and detection accuracy.
  • Limitations: Current models treat UAVs as a monolithic class, motivating fine-grained taxonomies that distinguish multi-rotors such as DJI Mavics from high-speed FPV drones.The paper identifies this distinction as essential for addressing tactical problems.
  • Limitations and future work: Processing audio in independent segments may limit detection stability, motivating sequential prediction integration to reduce false-positive rates.The proposed directions include memory modules or recurrent architectures for leveraging long-term dependencies in acoustic streams.
  • Limitations and future work: Lack of explicit spatial data, including distance and altitude, limits detailed error analysis and performance reporting.Future data collection should add distance metrics, UAV type, and other relevant metadata.
  • Future work: Future research should explore Transformer-based backbones, self-supervised pre-training, and semi-supervised pipelines for adapting to frontline locations with scarce labeled data.These directions target better use of large-scale unlabeled acoustic data and rapid adaptation to new locations.
Loading 2608.14287v1…