Overview
Whistle Classifier
Overview
The Whistle Classifier takes detected whistles from the Whistle and Moan detector and attempts to classify them to species.
The Whistle classifier does not attempt to classify individual detected whistles. Instead it collects information on many whistles and statistically analyses them as a group to determine species.
The classifier must be trained with a sample of whistles from each species. While some standard training sets are available through the PAMGuard web site, it is highly likely that whistles of a given species will vary between regions and sub species and may also evolve over time. The Whistle Classifier therefore allows you to train it using your own samples of whistles from known species.
Creating a Whistle Classifier
From the File>Add modules>Detectors menu, or from the pop-up menu on the data model display, select “Whistle Classifier”. Enter a descriptive name for the new detector and press OK.
General Principle of Operation
A major problem in the detection of whistles is that of overlapping whistles, or whistles which have been broken into parts during the detection process.
The two spectrograms in the picture show the same spectrogram data. The lower panel is overlaid with the output of the Whistle and Moan detector. Clearly some of the quieter whistles have been missed altogether, while others have been broken into multiple parts.
The breaking up of whistles into parts can be very dependent on the local noise conditions and would not therefore be consistent between encounters with a given species. A classifier requiring complete whistles would therefore be expected to perform very poorly.
The PAMGuard Whistle Classifier therefore works by taking short fragments of whistles and accumulates statistics over many fragments. Since the level of “natural” fragmentation due to noise will vary, all detected whistles are first broken into fragments of the same length prior to classification. While fragmenting whistles in this way is discarding potentially useful information, it creates an overall improvement in classifier stability
For each whistle fragment, three parameters are measured:
The start frequency
The slope
The curvature
Distributions of these three parameters are built up over time, then once sufficient fragments have been accumulated, the mean, the standard deviation and the skew (lopsidedness) of each distribution is measured, giving a total of nine parameters which are used by the statistical classifier, i.e.
The mean start frequency
The standard deviation of the distribution of start frequencies
The skew of the mean frequency distribution
The mean slope
etc. …
References
Gillespie, D., Caillat, M., Gordon, J., and White, P. (2013). �Automatic detection and classification of odontocete whistles,� The Journal of the Acoustical Society of America, 134, 2427�2437.