2004 ENGAGEMENT GRANT RECIPIENT
Tracking Humpback Whales with Passive Acoustics
Eva-Marie Nosal and L. Neil Frazer
Department of Geology and Geophysics, University of Hawaii at Manoa
Motivation: Since many marine mammals are very vocal, passive acoustic techniques present a highly promising approach to marine mammal studies, monitoring, and human impact mitigation. They can be used as an addition/alternative to more traditional visual and tagging techniques, which may be costly, invasive, and limited to favorable environmental considerations (such as sunlight and calm seas). Advantages of passive acoustic methods include relative cost-efficiency, non-invasiveness, and potential for long-term monitoring.
To track marine mammals using passive acoustics, vocalizations are recorded on several hydrophones (an array) towed behind a boat, mounted on the ocean-bottom, or suspended from buoys near the surface. These recordings are processed to find the position of the animal at incremental time steps. Our work involves the development and testing of the processing methods for tracking humpback whales.
Methods: The processing methods used depends on several factors, including: (1) the type of call being produced (including the duration of the call, its frequency range and characteristics - Fig. 1 shows a humpback whale signal); (2) environmental parameters (including bottom depth and sound-speed profile); and (3) the type of hydrophone array. Several techniques exist for localizing underwater sound sources, such as the “time difference of arrival” (TOAD) method and “matched-field” (MF) methods. There are several reasons that these methods are not always successful for tracking humpback whales. In TOAD methods, the longer-duration calls of humpbacks, the fact that they favor shallow waters, and the presence of multiple singers makes it difficult to identify and associate direct reflections. MF methods rely on accurate models of sound propagation (only possible for low frequencies), long line-arrays (costly), and knowledge of the source signature (not available for humpbacks).
To track humpbacks using arrays with a minimal number of hydrophones, we extended conventional MF techniques to get our “pair-wise spectrogram” (PWS) method. In this method, recordings are processed along pairs of receivers. The signal (data) at the first hydrophone in the pair is propagated (via convolution with a modeled Green's function) from the second hydrophone position to a candidate source location. This gives a resulting signal, H 12 . We get H 21 similarly. If the candidate source is at the correct source location, these two resulting signal waveforms should be identical. However, because we cannot model the environment perfectly and because there is noise in the data, they may differ significantly. Since spectrograms are less sensitive (than full waveforms) to noise and modeling uncertainties, we compare the spectrograms (rather than the waveforms) of H 12 and H 21 (by taking normalized inner products over frequency, time, and receiver pairs). This procedure is repeated for a grid of candidate source locations, and the one that gives the best agreement between spectrograms is the estimated source location.
Use of HPC resources: Computational demands associated with the prediction of Green's functions for each receiver/candidate source location pair increase with the product of frequency and range. Since our localization uses frequencies up to several kHz and we wish to localize at ranges of up to tens of kilometers, high-performance computing (HPC) resources are required to use the PWS method. Fortunately, the computations of Green's functions at different locations are independent, which means they parallelize well. Further computational demands are associated with spectrogram computation and inner products. Once again, since computations for difference candidate source locations are independent, this step parallelizes well. Once our algorithms were developed and implemented on a partial scale (fewer frequencies, shorter ranges) using desktop PCs, they were parallelized and ported to the Squall system at MHPCC for full scale testing and simulations.
Results and Future Work: We developed and tested PWS processing on simulated data . One of our more recent simulations is shown in Figure 1. The PWS method clearly outperforms both the TOAD and MF method in localizing both sources. We are currently collecting data of singing humpback in Hawaiian waters that will be used to test PWS in a real life situation. Processing of this data will require HPC resources for which we plan to continue to use the Squall system, and possibly the Tempest system, at MHPCC. We are also working on a variety of different processing techniques for localizing other species, such as sperm whales , blue whales, and spinner dolphins.
Figure 2 . Ambiguity surfaces (probabilistic indicators of source location with red indicating highest probability) in plan-view using the TOAD, MF, and PWS processing methods. 3 receivers were used to localize 2 sources (indicated with white triangles) in 0 dB signal-to-noise ratio. Environmental uncertainty is in the form of incorrect bottom depth: 200 m for the forward model, 204 m for the inversion. Only PWS finds both sources correctly.
Acknowledgments: Funding was provided by the University of Hawaii through a MHPCC Student Engagement Grant. HPC resources were provided by MHPCC.
 Nosal, E.-M., and Frazer, L.N., “Pair-wise processing of spectrograms for localization of multiple broadband CW sources,” Proceeding of the IEEE Oceans '05 Europe meeting, Brest, France, June 20-23 2005. Special publication in the Newsletter of the IEEE OES, Winter 2006. Nosal, E.-M., and Frazer, L. N., “Delays between direct and reflected arrivals used to track a single sperm whale,” Applied Acoustics, 87 (11-12), 1187-1201, 2006.