1
Unsupervised Classification of Marine Soundscape with Non-Negative Matrix Factorization Yuen-Wa HO, Tzu-Hao LIN, Tomonari AKAMATSU, Leszek KARCZMARSKI Email: [email protected]; Website: https://www.hku-cetacean-ecology.net/ METHODS INTRODUCTION Marine soundscape represents a combination of geophony, biophony and anthrophony in the marine environment (i.e. sound of geological processes, biological sounds and anthropogenic noise) Study species: Indo-Pacific humpback dolphin (Sousa chinensis) o Inhabits inshore coastal waters with depth < 20m o Highly susceptible to anthropogenic impacts This study aims to classify marine soundscape with nonnegative matrix factorization (NMF) with unsupervised approach to provide a quantitative acoustic measure of humpback dolphin habitats in Hong Kong waters RESULTS No prior labelling of the core habitat / non-core habitat is given in NMF decomposition (Unsupervised approach on spectral decomposition) New encoding matrix of the entire input data (600 recordings) was trained by using NMF (Fig 2) Decomposed spectral features were used to classify site-specific soundscape composition (Fig 3) (Supervised classification of marine soundscape) Spatial change of relative strength of each basis in different recording locations were constructed (Fig 4) DISCUSSION Coastal soundscape vary spatially in the habitat of Indo-Pacific humpback dolphins in Hong Kong waters Relative Strength of Spectral Characteristics corresponds with dolphin distribution data Information retrieval of sound-source contribution improves traditional approach delineating strength of spectral contribution along different frequency ranges Power Spectral Density (PSD) measurement by Discrete Fourier Transform (DFT) at 250 Hz interval between 250 Hz – 48kHz (Fig 1) Using PSD vector as the matrix for NMF decomposition Decomposition ran for 1000 times; Retaining significantly different basis between dolphin Core Habitat and Non-Core Habitat (p<0.001) Figure 1. Spectrogram of all recordings from Core Habitat and Non-Core Habitat Figure 2. NMF decomposition into 4 Spectral basis (Left Panel). Colours represent the relative contribution of that basis (y-axis) along each frame (x-axis) (Right Panel). ACKNOWLEDGEMENTS Ocean Park Conservation Foundation Hong Kong Cetacea Research Institute Data Collection: 3-min acoustic sampling at 15 field locations in Hong Kong waters across 2 years (2016-2018) Figure 3. Spectral features from 1000 trials of NMF decomposition (25%, 50%, 75% of spectral features) NMF Decomposition Result (50 %) as Prior Information(Labelled as Important Spectral Basis and Other Spectral Basis) for classifying site- specific marine soundscape composition with regard to dolphin distribution Figure 4. Relative Strength of Spectral Features in all recording sites

Unsupervised Classification of Marine Soundscape with Non ...Unsupervised Classification of Marine Soundscape with Non-Negative Matrix Factorization Yuen-WaHO, Tzu-Hao LIN, TomonariAKAMATSU,

  • Upload
    others

  • View
    6

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Unsupervised Classification of Marine Soundscape with Non ...Unsupervised Classification of Marine Soundscape with Non-Negative Matrix Factorization Yuen-WaHO, Tzu-Hao LIN, TomonariAKAMATSU,

Unsupervised Classification of Marine Soundscapewith Non-Negative Matrix Factorization

Yuen-Wa HO, Tzu-Hao LIN, Tomonari AKAMATSU, Leszek KARCZMARSKIEmail: [email protected]; Website: https://www.hku-cetacean-ecology.net/

METHODS

INTRODUCTION• Marine soundscape represents a combination of geophony, biophony and anthrophony in the marine

environment (i.e. sound of geological processes, biological sounds and anthropogenic noise)• Study species: Indo-Pacific humpback dolphin (Sousa chinensis)o Inhabits inshore coastal waters with depth < 20mo Highly susceptible to anthropogenic impacts

• This study aims to classify marine soundscape with nonnegative matrix factorization (NMF) with unsupervised approach to provide a quantitative acoustic measure of humpback dolphin habitats in Hong Kong waters

RESULTS• No prior labelling of the core habitat / non-core habitat is given in NMF decomposition (Unsupervised approach on spectral decomposition) • New encoding matrix of the entire input data (600 recordings) was trained by using NMF (Fig 2)• Decomposed spectral features were used to classify site-specific soundscape composition (Fig 3) (Supervised classification of marine soundscape)• Spatial change of relative strength of each basis in different recording locations were constructed (Fig 4)

DISCUSSION• Coastal soundscape vary spatially in the habitat of Indo-Pacific humpback dolphins in Hong Kong waters• Relative Strength of Spectral Characteristics corresponds with dolphin distribution data• Information retrieval of sound-source contribution improves traditional approach delineating strength of

spectral contribution along different frequency ranges

Power Spectral Density (PSD) measurement by Discrete Fourier

Transform (DFT) at 250 Hz interval between 250 Hz – 48kHz (Fig 1)

Using PSD vector as the matrix for NMF

decomposition

Decomposition ran for 1000 times; Retaining significantly different

basis between dolphin Core Habitat and Non-Core Habitat (p<0.001)

Figure 1. Spectrogram of all recordings from Core Habitat and Non-Core Habitat

Figure 2. NMF decomposition into 4 Spectral basis (Left Panel). Colours represent the relative contribution of that basis (y-axis) along each frame (x-axis) (Right Panel).

ACKNOWLEDGEMENTSOcean Park Conservation Foundation Hong KongCetacea Research Institute

Data Collection: 3-min acoustic sampling at 15 field

locations in Hong Kong waters across 2 years (2016-2018)

Figure 3. Spectral features from 1000 trials of NMF decomposition (25%, 50%, 75% of spectral features)

NMF Decomposition Result (50 %) as Prior Information(Labelled as

Important Spectral Basis and Other Spectral Basis) for classifying site-

specific marine soundscape composition with regard to dolphin

distribution

Figure 4. Relative Strength of Spectral Features in all recording sites