The Unbearable Weight: Scaling Models and Methods for UAV Audio Classification

1 Introduction

• Audio captures platform-specific rotor, motor, and aerodynamic signatures that can distinguish UAV models, while outdoor recordings also include environmental noise.Unlike telemetry or mission content, the recordings represent the drone itself.Source: §1 · p. 2
• Limited, costly-to-collect UAV audio makes training large networks from scratch impractical, while full fine-tuning is computationally heavy and slow to retrain.The interaction between model scale, parameter-efficient fine-tuning, and targeted augmentation on small-data audio tasks is not well-characterized.Source: §1 · p. 2
• The study compares CNN and transformer backbones at multiple sizes under full, classifier-only, and four parameter-efficient fine-tuning methods.It reports accuracy, training time, trainable-parameter share, and inference-time memory footprint on a 3,100-clip, 31-class dataset.Source: §1 · p. 2
• The central question is when large pretrained models’ compute and memory costs are justified versus when smaller models or lighter fine-tuning are better engineering choices.The introduction frames research questions around CNN-versus-transformer performance and accuracy-to-efficiency trade-offs under data scarcity.Source: §1 · p. 3

Previous Work

• The prior UAV-audio study used a 9-class dataset and found the custom CNN slightly outperformed the pretrained AST; the dataset later grew to 31 classes.The current dataset comprises 15,500 seconds of audio, enabling a look at scaling for UAV classification.Source: §2 · p. 3

UAV Classification

• Audio is presented as a practical, cost-effective UAV classification modality, though its short range and sensitivity to noisy environments constrain its use [1].The review contrasts audio with optical, radar, and radio-frequency sensing.Source: §2 · p. 4

Deep Learning & Audio Classification

• Prior work identifies CNNs as common in audio classification, while another comparison reports transformers can significantly outperform CNNs and supports transfer learning.The paper narrows its model selection to CNNs and transformers.Source: §2 · p. 4

Transfer Learning

• Transfer learning initializes a related task with weights from a model pretrained on a large dataset, and cross-modal transfer is an alternative when domain-specific pretrained networks are unavailable [8].The review describes it as reducing training and dataset-scarcity constraints.Source: §2 · p. 4

3.1 Data Collection

• 3,100 five-second recordings from 31 UAVs form a perfectly balanced dataset collected across indoor and outdoor environments in three U.S. locations.Recordings capture rotor, motor, and aerodynamic signatures during flight.Source: §3.1 · pp. 4–5

3.2 Technical Environment

• Experiments ran on a server with two RTX 4090 GPUs, but used one GPU because distributed runs took longer.The authors report code and training runs as publicly available.Source: §3.2 · p. 5

3.3 Feature Extraction

• Mel-spectrograms represent frequency content over time on a human-perception scale; extraction uses a 16,000 Hz sampling rate, 1,024-point FFT window, 128 hop length, and 128 mel bands.They are computed during training time.Source: §3.3 · p. 6

3.4 Data Augmentations

• Three deterministic compositions of time stretch and harmonic distortion are generated per training sample, tripling training data while leaving validation unaugmented.The authors note augmented data may not fully reflect real-world audio.Source: §3.4 · p. 7

3.5 Training

• Full fine-tuning updates all model parameters, whereas classifier fine-tuning trains only the final classifier block as an ablation of pretrained generalization.The paper describes classifier fine-tuning as not reliably large enough to generalize effectively.Source: §3.5 · pp. 7–8

3.6 PEFT

• SSF scales and shifts features across model families, IA3 and OFT target transformers, and batch-norm fine-tuning targets CNNs; the custom CNN is randomly initialized and evaluated only with full fine-tuning.Classifier fine-tuning is an ablation for the other backbones, while full fine-tuning is reported as an adaptation upper bound.Source: §3.6 · pp. 8–9

3.7 Models

• The model sweep includes a three-layer custom CNN, ResNet-18/152, MobileNet-V3 small/large, EfficientNet-B0/B7, ViT, and AST backbones.ViT uses an ImageNet-fine-tuned checkpoint, while the AST checkpoint was trained on AudioSet.Source: §3.7 · pp. 10, 11, 13

4 Experiments & Results

• All models used five-fold cross-validation and the same computational budget; training time, trainable-parameter share, and approximate memory footprint were tracked alongside accuracy.Each fold used an 80% training and 20% test split.Source: §4 · p. 13
• CNNs scale evenly with time to performance, whereas AST and ViT train longer with lower accuracy, making them less efficient.Training-time comparisons are shown in Fig. 8.Source: §4 · p. 14, Figure 8 · p. 16
• 97.65%±0.30 validation accuracy was achieved by EfficientNet-B7 with batch-norm fine-tuning and three-fold augmentation.Table 1 reports the five-fold cross-validation results.Source: §4 · p. 13, Table 1 · p. 15

RQ1: CNNs vs. transformers

• 97.65% ± 0.30 for EfficientNet-B7 exceeds AST with OFT at 96.74%±0.24; the CNN trains in roughly one-seventh the wall time per fold and uses 243 MB versus 333 MB.Source: §5 · p. 16

RQ2: best accuracy-to-efficiency configuration

• 97.65% ± 0.30 validation accuracy, 0.49% of model parameters trained, and a 243 MB inference-time memory footprint make EfficientNet-B7 with batch-norm fine-tuning and three-fold augmentation the strongest configuration.The configuration updates fewer than 0.5% of parameters.Source: §5 · pp. 16–17

RQ3: within-family backbone scaling

• Larger backbones make PEFT effective across families, while smaller backbones require full fine-tuning to remain competitive; MobileNetV3 is an exception to the scaling pattern.EfficientNet-B7 peaks under batch-norm fine-tuning, while EfficientNet-B0 peaks under full fine-tuning.Source: §5 · p. 17

Why CNNs outperform transformers

• The discussion attributes CNN performance to locality suited to mel-spectrograms, weak alignment between inherited pretraining and UAV audio, and batch-norm adaptation of channel statistics.The reported custom CNN result is 94.32% ± 1.09% with three-fold augmentation, illustrating the boundary case without a strong pretrained signal.Source: §5 · p. 17

Future directions

• Future work targets continual adaptation, multimodal fallback, larger-scale transformer crossover tests, and separating feature transfer from distribution-level adaptation.The paper says these questions remain open, including how PEFT methods depend on data scale.Source: §5 · p. 18

Conclusion

• For UAV audio classification under data scarcity, the conclusion favors lightweight CNN backbones with selective fine-tuning over large transformers with full fine-tuning.The paper characterizes scaling the method, rather than the model, as giving the best balance of accuracy, training efficiency, and inference-time footprint.Source: §5 · p. 18

• The work is supported by the South Carolina Research Authority (SCRA).Source: §6 · p. 18
• The code and model training logs are available online, and the authors declare no conflict of interest.Source: §6 · p. 18
• The authors state that ethical approval was not applicable.Source: §6 · p. 18

• Table A1 is described as listing the 31 UAV audio dataset classes with collection sites.The appendix description also names collection site, manufacturer, and model name as table fields.Source: Appendix A · p. 22, Table A1 · p. 23

Appendix B Memory Footprint

• The reported memory footprint is an inference-time measure for each model.The paper identifies it as relevant to edge devices and overall scale.Source: Appendix B · p. 24
• The memory estimate uses bytes per parameter; for float32, the paper specifies four bytes per parameter.Source: Appendix B · p. 24
• Model accuracy and memory footprint show little correlation, and additive PEFT methods such as SSF, IA3, and OFT slightly raise the footprint.Fig. B1 illustrates the comparison, and Table B2 provides the model and PEFT breakdown.Source: Appendix B · p. 24, Figure B1 · p. 24

Appendix C Training Time vs. Augmentations

• Data augmentations significantly increase training times.Source: Appendix C · p. 24

Appendix D Training Time Table

• Table D3 reports average training runtime in minutes for different models under various PEFT configurations.Source: Appendix D · p. 24
• Blank entries in Table D3 indicate PEFT incompatibility.Source: Appendix D · p. 24

Appendix E Trainable Parameter Percentage Table

• Full fine-tuning trains all parameters, corresponding to 100%.Source: Appendix E · p. 25
• Table E4 reports trainable-parameter percentages for each model and PEFT configuration.Source: Appendix E · p. 25, Table E4 · p. 27