Page 43 - Acoustics Today
P. 43

                                 Trials with known-location sounds—harbor porpoise in Denmark
Yet another method of obtaining the detection probability
multiplier is possible when you have a set of animals at known
locations relative to the hydrophones. You can then record, for
each of a set of time periods, whether each animal was detect-
ed by each hydrophone, and use these data to construct a
regression of probability of detection against distance (and
perhaps other factors). This is the approach taken in our first
example, above, where the beaked whales were tagged in the
vicinity of the AUTEC hydrophones, and hence their location
each time they clicked could be derived, as well as whether
each of these clicks were detected on the surrounding
hydrophones. Another example of this kind of method is the
study of Kyhn et al. (2012), who were interested in determin-
ing the feasibility of estimating the density of harbor porpoise
Acoustic modeling—beaked whales in the Bahamas, again
The above methods all use additional data, hopefully col- lected at the time of the survey, to estimate the detection probability multiplier. However, in some cases such data are not available—all we know is how many sounds were detect- ed on each sensor. In this case, it may still be possible to esti- mate detectability, if we are willing to make some strong assumptions about sound production, propagation, and pos- sibly detection. This involves application of the passive sonar equation, which can be written
SNR = SL - DL - TL - NL (5)
where SNR is the signal to noise ratio at the receiver, SL is source level, DL is the directivity loss (i.e., attenuation in the source level due to the direction the animal is pointing rela- tive to the hydrophone when they make the sound), TL is transmission loss as a function of the distance between ani- mal and receiver and NL is the level of ambient noise. Hence, given a set of assumptions about the distribution these quan- tities, one can predict the distribution of SNR for a sound produced at a given location. SNR can be related to detection probability either by assuming that sounds above a certain SNR are certain to be detected, or (better) using an empiri- cally-derived relationship between probability of detection and SNR.
An example of this is given by Küsel et al. (2011), who derived a density estimate for Blainville’s beaked whale based on data from a single hydrophone at the AUTEC range. They obtained distributions of SL, DL, and NL from the published literature, and used several models of sound propagation to calculate TL. The relationship between detection and SNR
(Phocoena phocoena) using commercially-available 6
autonomous porpoise detectors called T-PODs. Kyhn et al. (2012) set up a visual monitoring station on cliffs at Fyns Hoved, Denmark, and tracked passing porpoises. They moored T-PODs below the cliffs, and by splitting each track up into a set of short segments, could determine for each segment whether the T-POD registered or not the presence of the por- poise.
 was estimated using a short sample of data from the study period that was hand-annotated and then run through the detector. From these pieces, average detection probability was derived using a Monte Carlo procedure, sampling from assumed distributions of animal position and orientation, SL, DL, NL and estimated values of the SNR-detection probabil- ity regression parameters. The resulting estimate, 0.014 (CV 17.9%) was not too dissimilar from that obtained by Marques et al. (2009) using the DTAG data (0.032 with CV 15.9%). Clearly, however, measurements will be preferable to assumptions whenever possible.
A half-way-house between this and the distance-based approach discussed above would be to use the received level of calls to estimate their distance, by using an assumed trans- mission loss model and source level distribution, or a regres- sion relationship built using a sample of sounds where received level and actual distance are known. One example of estimating distance from received level is Širović et al. (2009), using blue whales in the Antarctic Ocean.
Conclusions
Those pesky multipliers
We have shown that it is possible to estimate animal density from passive acoustic data from fixed sensors, and have demonstrated a variety of approaches. The best method depends on the type of data available, and what you can most reliably count. If you can count all of the individ- uals within the study area and exclude all of those outside, then reliable estimation is straightforward. In most situa- tions, however, this is not possible, and you then need mul- tipliers to convert the count from calls or groups into ani- mals, and to deal with false positive and false negative rates. We have demonstrated several methods for estimating detectability (the complement of false negative rate). The most reliable use data collected at the same time as the main survey, such as distances to detections, or associations in detections among sensors. A less-satisfactory option is to undertake a secondary survey, such as tagging some ani- mals—again it is best if this is done at the same time and place as the main survey but this is often not possible. Deriving detection probability estimates from acoustic modeling alone is a last resort, but often constraints will mean that other approaches are not possible.
Often the object being counted is an individual vocaliza- tion: an echolocation click or a call. We then need an estimate of the vocalization rate. This is often the main impediment to obtaining a density estimate—either the rate is completely unknown, or it has been measured in a very different cir- cumstance and could reasonably be expected to vary over time or space. Knowledge of the basic acoustic biology of our study species is often a fundamental limitation.
Other applications
We have focused here on fixed detectors, but there is plenty of potential for estimating density from towed acoustic sensors, behind ships or gliders for example. If the sensors are designed to get the bearing to detected calls, and animals vocalize repeatedly, then intersecting bearings can be
42 Acoustics Today, July 2012




































































   41   42   43   44   45