Classification of Bryde's whale individuals using high-resolution time-frequency transforms and support vector machines.
Journal:
The Journal of the Acoustical Society of America
PMID:
40130953
Abstract
Whales generate vocalizations which may, deliberately or not, encode caller identity cues. In this study, we analyze calls produced by Bryde's whales and recorded by ocean-bottom arrays of hydrophones deployed close to the Costa Rica Rift in the Panama Basin. These repetitive calls, consisting of two main frequency components at ∼20 and ∼36 Hz, have been shown to follow five coherent spatiotemporal tracks. Here, we use a high-resolution time-frequency transform, the fourth-order Fourier synchrosqueezing transform, to extract time-frequency characteristics (ridges) from each call to appraise their suitability for identifying individuals from each other. Focusing on high-quality calls recorded less than 5 km from their source, we then cluster these ridges using a support vector machine model resulting in an average cross-validation error of ∼11% and balanced accuracy of ∼86 ± 5%. Comparing these results with those obtained using the standard short-time Fourier transform, k-means clustering, and lower-quality signals, the Fourier synchrosqueezing transform approach, coupled with support vector machines, substantially improves classification. Consequently, the Bryde's whale calls potentially contain individual-specific information, suggesting that individuals can be studied using ocean-bottom data.