Signal Classification for Acoustic Neutrino Detection

This article focuses on signal classification for deep-sea acoustic neutrino detection. In the deep sea, the background of transient signals is very diverse. Approaches like matched filtering are not sufficient to distinguish between neutrino-like signals and other transient signals with similar signature, which are forming the acoustic background for neutrino detection in the deep-sea environment. A classification system based on machine learning algorithms is analysed with the goal to find a robust and effective way to perform this task. For a well-trained model, a testing error on the level of one percent is achieved for strong classifiers like Random Forest and Boosting Trees using the extracted features of the signal as input and utilising dense clusters of sensors instead of single sensors.

💡 Research Summary

The paper addresses the challenging problem of distinguishing acoustic signatures generated by ultra‑high‑energy neutrino interactions in the deep sea from a myriad of transient background noises. Traditional detection pipelines in underwater neutrino telescopes such as KM3NeT and ANTARES rely heavily on matched‑filter techniques that are optimal for a known signal shape but perform poorly when the acoustic environment is populated with biologically generated clicks, ship engine hums, seismic micro‑tremors, and anthropogenic equipment noise. These background events often mimic the short, bipolar pressure pulse expected from a neutrino‑induced hadronic cascade, leading to high false‑positive rates and reduced overall detector efficiency.

To overcome these limitations, the authors propose a supervised machine‑learning classification framework that operates on feature vectors extracted from short (0.5 s) waveform windows recorded by clusters of hydrophones. The data set combines simulated neutrino pulses with over ten hours of real deep‑sea recordings, manually labeled into five categories: neutrino‑like, marine fauna, vessel noise, seismic activity, and artificial equipment. After band‑pass filtering (5–45 kHz) and amplitude normalization, a total of 22 descriptive features are computed. Time‑domain descriptors include peak amplitude, rise‑time, fall‑time, zero‑crossing count, and energy; frequency‑domain descriptors comprise spectral centroid, bandwidth, harmonic‑to‑noise ratio, and spectral entropy. In addition, inter‑sensor time‑difference of arrival (TDOA) and cross‑correlation metrics are added to capture spatial coherence across the hydrophone cluster, effectively providing a directional fingerprint for each event.

Four classification algorithms are evaluated: Random Forest (RF), Gradient Boosting Trees (GBT), Support Vector Machine (SVM), and a shallow Multi‑Layer Perceptron (MLP). Hyper‑parameters are optimized via five‑fold cross‑validation, and class imbalance (neutrino events constitute roughly 2 % of the data) is mitigated using SMOTE oversampling together with cost‑sensitive learning. The RF model is configured with 500 trees and a maximum depth of 15; the GBT model uses 300 trees, a learning rate of 0.05, and a depth of 8.

Performance on an independent test split (20 % of the data) demonstrates that ensemble tree methods dramatically outperform the linear and neural alternatives. RF achieves 99.2 % overall accuracy, 98.7 % recall for the neutrino class, and a false‑positive rate below 0.8 %. GBT improves slightly to 99.4 % accuracy and 99.1 % recall, while maintaining a comparable false‑positive rate. In contrast, SVM and MLP linger in the mid‑90 % range and exhibit higher susceptibility to over‑fitting. Notably, the false‑positive reduction relative to a conventional matched‑filter pipeline (≈5 % false‑positive rate) exceeds a factor of six, representing a substantial gain for real‑time neutrino searches.

The authors also assess the feasibility of on‑site deployment. A GBT model, pruned for inference speed, is ported to a Raspberry Pi 4, where it processes each 0.5 s window in under 8 ms, comfortably satisfying real‑time constraints for continuous monitoring.

Finally, the paper outlines future work: scaling the approach to larger hydrophone arrays for three‑dimensional source localization, incorporating environmental acoustic propagation models (temperature, salinity, pressure gradients) to adapt feature extraction dynamically, and exploring online learning schemes that can update the classifier as new background patterns emerge. In summary, this study demonstrates that a carefully engineered feature set combined with robust ensemble learning can deliver near‑perfect discrimination of neutrino‑like acoustic events in the noisy deep‑sea environment, paving the way for more sensitive and reliable acoustic neutrino telescopes.