Classifying Fish Sounds Using Wavelets
Mark Wood
Biomathematics and Statistics Scotland, Rowett Research Institute, Greenburn Rd., Aberdeen, AB21 9SB, Scotland, UK mark@bioss.ac.uk

Introduction

Many species of marine fish emit low frequency sounds composed of sequences of nearly identical transient units. The production of these sounds is often coupled with displays of aggression and/or courtship. In order to associate sound production with fish behaviour we need to be able to distinguish between the sounds of different species and between individual fish to be able to identify which fish is emitting sound at any given time.

Wavelets have been used to produce features of the waveforms which are then used to discriminate between the sounds from different fish. We consider the performance of this method for discriminating between individual haddock, Melanogrammus aeglefinus, and for discriminating between sounds from three fish species, the haddock, cod Gadus morhua, and pollack Pollachius pollachius.

Data

Recordings of the haddock were made at the FRS Marine Laboratory, Aberdeen in a semi-annular tank (90m³) containing 3 male and 5 female fish. The sounds were detected by a broad-band hydrophone, amplified and sampled at a frequency of 8 kHz. The haddock were maintained under controlled conditions, and recordings were made over two spawning seasons (February-April, 1999 and 2000). Haddock sounds consisted of long trains of regularly repeated ‘knocks’.

Figure 1 shows a typical recording consisting of a series of regularly spaced low frequency sound units, or knocks. Figure 2 shows the different waveforms of the 3 male haddock. The haddock varied their sound by repeating these knocks at different rates.

The cod sounds were recorded in the aquarium of the FRS Marine Laboratory and consisted of long grunts, produced singly or in groups of up to 5. The pollack sounds were recorded in the sea at a depth of 15m in Loch Torridon, Wester Ross, from a cage of fish, and consisted of short repeated grunts. The sounds of all three species are described by Hawkins and Rasmussen (1978).

Figure 1 - Sound recording of a single haddock made in the tank.

Figure 2 - The waveform produced by the haddock.

Wavelets

Wavelets are special mathematical functions, designed to overcome the shortfalls of the well known Fourier Transform. Wavelets are produced by scaling (compressing or expanding) and shifting a single ‘mother’ wavelet along the time axis. These wavelets are usually designed to form an orthonormal basis, in which any sound signal may be represented as a series of the scaled and shifted wavelets. The ‘amount’ of each wavelet present in the decomposition determines the dominant frequency components and their location in the signal. For this reason we say that the wavelet transform has good time and frequency localization.

Background material on wavelet analysis may be found in Jawerth and Sweldens (1994), Bruce and Gao (1996) and Abramovich et al. (2000). A more mathematical treatment is given by Chui (1992a,b) and Daubechies (1992).

Recognition of Individual Haddock

Wavelets were used to extract features from the sound units which would enable individual haddock or different species to be automatically recognised. The procedure consisted of 4 steps.

1. The smoothing property of wavelets was used to automatically isolate individual sound units in the recordings. By setting certain smaller wavelet coefficients to zero and then applying the inverse wavelet transform we were able to extract individual haddock knocks, or cod and pollack grunts from the background. These were extracted in 32ms windows (containing 256 data points) and were standardised so that their amplitude was of unit variance.
2. Due to the way in which the knocks were extracted in (1), and the fact that the wavelet transform is sensitive to shifts along the time scale, the non-decimated (or stationary) wavelet transform was used to decompose the knocks. A member of the Coiflet family of wavelets was found to give the best results.
3. Plots of the non-decimated wavelet coefficients in descending order of absolute value in each level are very similar for sounds from the same fish, but clearly different for sounds from different fish. These plots suggest suitable features for discriminating between the knocks of individual male haddock and between different fish species. For haddock, a plot of the scores on the first two canonical variates (Krazanowski (1996)) showed three well separated clusters.
4. Certain features from (3) were selected and used in a discriminant analysis to allocate unclassified knocks to one of the three male haddock or to each of the three species, or to a spurious sound. The method of extracting pulses in (1) meant that sounds were picked up which were not produced by any of the male haddock, cod or pollack. These spurious sounds, caused by splashes for example, could not be eliminated and so were allocated to a fourth group.

Results and Conclusions

In a test data set of the haddock sounds, 175 knocks were detected as having come from fish A, 336 from B, 194 from C and 142 were spurious sounds. The classification rates are shown in table 1. The overall success rate was 89%. It was shown that using the fact that knocks occurred in long repetitive series increased the success rate to 95%.

For the allocation of sounds to different species, 5 sounds were detected as having come from cod, 609 from haddock, 151 from pollack and 94 of the sounds were spurious. The classification rates are shown in table 2. The overall success rate achieved was 83%.

Table 1 - Classification rates of individual haddock knocks.

Table 2 - Classification rates of cod/haddock/pollack sounds

Wavelets provide a useful method of automatic sound recognition. The methods described above can count and assign a large number of sounds far more quickly than can be done by eye. This technique has the potential to separate fish sounds from ambient noise in the sea, and may provide a non-invasive method for locating spawning fish.

Acknowledgements

We thank Professors I.T. Jolliffe and A.D. Hawkins, University of Aberdeen, and Dr G. Horgan, Biomathematics and Statistics Scotland, for their supervision and input. We thank Licia Casaretto, FRS Marine Laboratory, for recording the haddock sounds, and Professor A.D. Hawkins for supplying the cod and pollack recordings. We acknowledge the financial support provided by the sponsors of the International Workshop on the Application of Passive Acoustics in Fisheries, Dedham, Boston MA.

References

Return to Top | Workshop Proceedings: Short Papers