Lightweight Underwater Acoustic Vessel Classification for Embedded Deployment

Abishek Soti; Thura Pyae Sone; Naqib Ibnul; Jim Huang; Sergio Chevtchenko; Henry Zhong; Gregory Cohen; Ying Xu

cs.SDOct 29, 2026

Lightweight Underwater Acoustic Vessel Classification for Embedded Deployment

Authors: Abishek Soti, Thura Pyae Sone, Naqib Ibnul, Jim Huang, Sergio Chevtchenko, Henry Zhong, Gregory Cohen, Ying Xu

Organizations: ICNS, Western Sydney University, Sydney, Australia

Abstract

We propose a Raspberry Pi-based underwater vessel classification system for edge deployment under limited connectivity, bandwidth and power in offshore acoustic surveillance. Using ShipsEar, we compare six classifiers on a Raspberry Pi 5 through a complete audio-to-prediction pipeline. We assess classification performance, processing latency, CPU utilisation, memory consumption and monitored power. Supporting experiments examine acoustic feature selection, statistical pooling and Transformer patch configurations. The two-layer convolutional neural network (CNN) achieved a mean macro-F1 of (99.14 \textpm{} 0.32)% across five random seeds. The radial basis function support vector machine (RBF-SVM) achieved 98.83%, with a median end-to-end latency of 28.109~~ms compared with 120.655~~ms for the CNN. This favourable trade-off motivated RBF-SVM deployment on the Raspberry Pi Zero 2 W. It maintained the same classification performance with a median end-to-end latency of 349.09~ms and a sustained throughput of 2.962 classifications/s, or 14.8 times the rate required for consecutive five-second audio segments.

Figures & tables

Fig. 1: Raspberry Pi 5 underwater vessel classification pipeline. Five-second audio segments are converted into a

128\times 475

STFT representation. The CNN and Patch Transformer use spectrograms; the other four classifiers use 256-dimensional mean and standard deviation features.

CNN: Statistical pooling
Configuration	Input / Patches	Params	Macro-F1 (%)
Without pooling	302,080	$\sim$ 1,668,000	89.87
Mean + std	5,120	183,237	96.98
Transformer: Patch geometry
Local ( $8\times 20$ )	368	201,349	74.53
Full freq. ( $128\times 20$ )	23	464,389	96.85

TABLE I: Effect of CNN statistical pooling and Transformer patch geometry on ShipsEar classification.

Model	Acc. (%)	Macro-F1 (%)	Size (KiB)	Infer. (ms)	E2E (ms)
CNN (2-layer)	$\mathbf{99.06\pm 0.33}$	$\mathbf{99.14\pm 0.32}$	723.08	91.050	120.655
Patch Transformer	$96.40\pm 0.71$	$96.48\pm 0.63$	1824.11	2.776	29.738
MLP	$96.05\pm 0.79$	$96.08\pm 0.71$	1179.03	0.901	27.772
Linear SVM	$90.34\pm 0.00$	$89.70\pm 0.00$	17.25	0.614	27.179
Logistic regression	$87.87\pm 0.00$	$88.03\pm 0.00$	12.39	0.652	27.270
RBF-SVM †	$98.88\pm 0.00$	$98.83\pm 0.00$	2428.27	1.193	28.109

TABLE II: ShipsEar classification performance and latency on Raspberry Pi 5.

Model	CPU (%)	RAM (MiB)	PMIC (W)	Max. temp. ( ∘ C)
CNN (2-layer)	54.56	490.52	5.193	54.50
Patch Transformer	30.48	452.11	4.153	50.10
MLP	25.16	287.25	3.960	50.15
Linear SVM	25.15	291.19	3.895	52.35
Logistic regression	25.56	290.69	4.014	52.35
RBF-SVM	25.16	293.14	3.863	48.50

TABLE III: Hardware resource utilisation on Raspberry Pi 5.

Fig. 2: Classification accuracy versus median end-to-end latency for six classifiers on the Raspberry Pi 5. The RBF-SVM achieved 98.88% accuracy compared with 99.06% for the CNN, while providing a 4.29-fold speedup.

Metric	Pi 5	Pi Zero 2 W
Feature extraction (ms)	26.10	327.88
SVM inference (ms)	1.19	9.40
Peak process RAM (MB)	307.4	119.02
RAM / device capacity (%)	7.7	24.4
Mean process CPU (%)	25.07	25.25

TABLE IV: RBF-SVM processing performance and resource usage on Raspberry Pi 5 and Zero 2 W using ShipsEar.

Audio Classification

Audio classification assigns recordings to categories such as speech, music, animal calls, acoustic scenes, and health-related events.

See all 149 papers

Towards Deployable Underwater Vessel Classification
Sep 24, 2026Abishek Soti, Thura Pyae Sone, Naqib Ibnul +4Passive Acoustic Monitoring Lightweight CNNs
ICNS, Western Sydney University, Sydney, Australia
Hearing the Ocean: Bio-inspired Gammatone-CNN framework for Robust Underwater Acoustic Target Classification
May 6, 2026Rajeshwar Tripathi, Sandeep Kumar, Monika Aggarwal +1Convolutional Neural Networks Audio Classification
Central Research Laboratory, Bharat Electronics Limited, Ghaziabad, India · Centre for Applied Research in Electronics (CARE), IIT Delhi, India
Linear probing enables Ship-Radiated Noise recognition with pretrained audio embeddings
Jan 13, 2026Hilde I. Hummel, Sandjai Bhulai, Rob D. van der Mei +1Audio Representation Learning Linear Probing
Centre of Mathematics and Computer Science, Stochastics group, Science Park 123, Amsterdam, 1098 XG, Netherlands · Vrije Universiteit, Department of Mathematics, De Boelelaan 1111, Amsterdam, 1081 HV, Netherlands · Naturalis Biodiversity Center, Darwinweg 2, Leiden, 2333 CR, Netherlands

Sep 24, 2026J/K move · Enter open · S save

Towards Deployable Underwater Vessel Classification

Abishek Soti, Thura Pyae Sone, Naqib Ibnul, Htoo Htet Aung+3

ICNS, Western Sydney University, Sydney, Australia

Passive Acoustic Monitoring Lightweight CNNs Audio Classification Underwater Acoustics

Open paper AskPDF arXiv

We propose a compact underwater acoustic classification framework combining multi-representation feature engineering, temporal statistical pooling, and compact convolutional architectures designed for acoustic time-frequency and cochlear representations. We investigate multiple conventional and auditory-inspired representations and first evaluate lightweight classifiers and Conventional Neural Networks (CNNs) on ShipsEar dataset. On the provided split, a two-layer CNN achieves a macro F1 of 0.9918, while a Radial Basis Function Support Vector Machine (RBF-SVM) reaches 0.9883. However, source-recording provenance cannot be reconstructed, preventing verification of recording-independent generalisation. We therefore evaluate on DeepShip dataset using recording-level partitioning before segmentation. Under this protocol, a 157K-parameter compact CNN achieves a test macro F1 of 0.7226, while an 11.17M-parameter ResNet18 provides no improvement in validation performance under the matched setting. These results demonstrate the importance of representation-aware feature and model design, together with rigorous recording-level evaluation, for classification performance and deployability in compact underwater acoustic systems.

Abstract

Audio Classification

Explore similar work

Towards Deployable Underwater Vessel Classification

Hearing the Ocean: Bio-inspired Gammatone-CNN framework for Robust Underwater Acoustic Target Classification

Linear probing enables Ship-Radiated Noise recognition with pretrained audio embeddings