15,500 Seconds: Lean UAV Classification Leveraging Parameter Efficient Fine-Tuning and Pre-Trained Networks
Andrew P. Berg, Qian Zhang, Mia Y. Wang · 2025
Read 15,500 Seconds: Lean UAV Classification Leveraging Parameter Efficient Fine-Tuning and Pre-Trained Networks by Andrew P. Berg, Qian Zhang, Mia Y. Wang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-11049.txt.
Introduction
Andrew Berg, Qian Zhang, and Mia Wang sort drone types by sound with a lean training recipe, in a 2025 paper. Their set holds 3,100 audio clips, 15,500 seconds in all, across thirty-one drone types. Data that small starves big models, so the team borrows pre-trained networks and tunes them lightly.
They compared a custom small network, a transformer, and EfficientNet under varied tuning plans and sound tweaks. Full tuning of EfficientNet with three tweaks won, reaching about ninety-six percent accuracy in cross-checks. Light tuning plus smart tweaks, they conclude, beats bulk when drone sound data are scarce. Future work will add sight and radar feeds.
Catalog record
| CREATOR | Andrew P. Berg, Qian Zhang, Mia Y. Wang |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 24 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-11049
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
15,500 Seconds: Lean UAV Classification Leveraging Parameter Efficient Fine-Tuning and Pre-Trained Networks Andrew P. Berg, Qian Zhang, Mia Y. Wang Abstract Unmanned Aerial Vehicles (UAVs) pose an escalating security concerns as the market for consumer and military UAVs grows. This paper address the critical data scarcity challenges in deep UAV audio classification. We build upon our previous work expanding approaches such as: parameter efficient fine-tuning, data augmentation, and pre-trained networks. We achieve performance upwards of 95% validation accuracy with EfficientNet-B0. I Introduction Unmanned Aerial Vehicles (UAVs), or commonly known as drones, pose security and intelligence threats to private and public interests. There is a growing need in defense and intelligence sectors for UAV Classification as modern methods and systems rely on less-robust UAV detection. UAV detection is not enough, as more drones flood the airs. There are a variety of ways to classify UAVs using various modalities: Radio frequency (RF), Radar, visual, and audio. In this work we focus on UAV audio classification as a starting point in future works we aim to improve the scope incorporating multiple modalities. Further, because of the scale and complexity of our data we have opted to use deep learning (DL) neural networks. A key issue in all of the listed modalities is data scarcity. It is relatively expensive to capture the necessary telemetry for all available UAVs to classify. In this paper we build on our lab’s previous work to solve this problem using data, and architecture training regimes such as: parameter efficient fine-tuning, data augmentation, and pre-trained networks. In the following sections we will discuss previous work and literature that encouraged our findings. A deep dive into our methodology, including: how our dataset was collected, training environment, and a breakdown of the solutions used to train for classification networks. Lastly, we will provide an empirical analysis with 5-fold cross validated results, where we achieve upwards of 95% validation accuracy using EfficientNet. The projects research questions are as follows: 1. On our custom UAV dataset can fine-tuned transformers outperform fine-tuned CNNs? 2. What is the optimal combination of PEFT, data augmentations, and model selection to achieve the highest result? II Literature Review Previous Work: This paper is primarily based on the previous work done by our lab in [ 1 ] . The prior paper explored an extremely small data approach of a 9-class dataset (4,500 seconds) and UAV audio classification; comparing transformers and CNNs. The former paper concluded that our custom CNN approach had better performance. Importantly, when working with deep learning models such as CNNs and transformers what we deem small data is entirely relative to the model architecture. Our custom dataset is large and complex enough for a less complex machine learning (ML) system, whereas the neural networks are more data hungry. Audio Classification: We use deep audio classification to refer to audio classification performed using neural networks. The study [ 2 ] . Found the most common approach to involve CNNs, using feature extraction techniques to transform the raw audio content using various methods like relative waves and mel-scale transformations. Another paper [ 3 ] compares pre-trained popular transformers and CNNs against one another further motivating our work, and guiding our model selections. UAV Classification: Reference [ 48 ] breaks down the various modalities and their drawbacks. Further it explores the different machine and deep learning approaches to the problem. Other studies look into the various open UAV datasets [ 48 ] and approaches and the different approaches taken to them. III Methodology III-A Data Collection The drone data collection is still an on-going project [ 50 ] [ 51 ] . So far, for the past 5 years, drone audio data was collected from 31 distinct unmanned aerial vehicles (UAVs), as detailed in Table I . Each UAV contributed 100 five-second audio recordings, resulting in a total of 3,300 audio files and 15,500 seconds of raw flight audio. The dataset is perfectly balanced. These recordings encompass a broad spectrum of consumer, commercial, and custom-built UAV platforms and were conducted across both indoor and outdoor environments in three U.S. locations. Drone Overview: The dataset includes 28 quadcopters, one tricopter, and one hexacopter. Most of the quadcopters use a conventional X-frame design typical of consumer drone platforms. UAVs were sourced from manufacturers including DJI, Autel, Syma, Yuneec, UDI, Hasakee, Holystone, Hover, and two self-built drones. Notably, David Tricopter , designed by David Windestal, is a custom-built tricopter with a 34-inch diameter, equipped with an AfroFlight Naze32 flight controller and weighing approximately 2.6 lbs. PhenoBee , designed by Ziling Chen, is the largest UAV in the dataset, weighing approximately 23 kg with a 1.35-meter diameter, and operates using the Ardupilot framework with Cube Orange hardware. Recording Sites: Data were collected in West Lafayette, Indiana; New Richmond, Indiana; and Charleston, South Carolina. Indoor recordings in Indiana were conducted in a university lab setting, while outdoor recordings were collected on a private farm in New Richmond. Charleston indoor data were collected in the Drone Lab at the Harbor Walk Campus of the College of Charleston, and outdoor recordings were captured from the rooftop of the aquarium parking garage. Environmental factors varied naturally during recordings and included wind, birdsong, traffic noise, and changing weather conditions. Recording Equipment: From 2021 through 2023, audio data were captured using a MacBook Air with a 1.1GHz quad-core Intel Core i5 processor and 8GB of memory. Beginning in 2024, recordings used a MacBook Air with an Apple M3 chip and 16GB of memory. All data were recorded using the system’s built-in microphones without external post-processing. TABLE I: UAV Audio Dataset: 31 Classes with Collection Sites Manufacture Model Drone Type Number of Files Duration (sec) Collection Site Self-build David Tricopter Outdoor 100 500 Columbus, IN Self-build PhenoBee Outdoor 100 500 West Lafayette, IN Autel Evo 2 Pro Outdoor 100 500 New Richmond, IN DJI Avata Outdoor 100 500 Charleston, SC DJI FPV Outdoor 100 500 Charleston, SC DJI Matrice 200 Outdoor 100 500 West Lafayette, IN DJI Matrice 200 V2 Outdoor 100 500 New Richmond, IN DJI Matrice 600p Outdoor 100 500 New Richmond, IN DJI Mavic Air 2 Outdoor 100 500 New Richmond, IN DJI Mavic Mini 1 Outdoor 100 500 New Richmond, IN DJI Mini 2 Outdoor 100 500 New Richmond, IN DJI Mini 3 Outdoor 100 500 Charleston, SC DJI Mini 3 Pro Outdoor 100 500 Charleston, SC DJI Mavic 2 Pro Outdoor 100 500 New Richmond, IN DJI Mavic 2s Outdoor 100 500 New Richmond, IN DJI Phantom 2 Outdoor 100 500 New Richmond, IN DJI Phantom 4 Outdoor 100 500 New Richmond, IN DJI Tello Indoor 100 500 Charleston, SC DJI RoboMaster TT Tello Indoor 100 500 New Richmond, IN Hasakee Q11 Indoor 100 500 West Lafayette, IN Holystone HS210 Indoor 100 500 Charleston, SC Hover X1 Outdoor 100 500 Charleston, SC Syma X5SW Indoor 100 500 West Lafayette, IN Syma X5UW Indoor 100 500 West Lafayette, IN Syma X8SW Indoor 100 500 West Lafayette, IN Syma X20 Indoor 100 500 West Lafayette, IN Syma X20P Indoor 100 500 West Lafayette, IN Syma X26 Indoor 100 500 West Lafayette, IN Swellpro Splash 3 plus Outdoor 100 500 New Richmond, IN Yuneec Typhoon H Plus Outdoor