Cetacean Monitoring Systems
The design of C-PODs aims to solve four key problems in monitoring cetacean echo-location:
Origin and design
In 1991-2 volunteer observers on the gill netting fleets of Cornwall, UK, and southern Ireland found a large bycatch of porpoises (Phocoena phocoena) where a large decline in small cetaceans was already known. To provide a constructive way forwards a tool was needed to reveal the porpoises' movements around the nets. The first POD (POrpoise Detector), retrospectively called the 'Proto-POD', was developed to meet this need.
It was based on a system for detecting porpoise clicks from a moving yacht that had already been developed by Oliver Chappell, Russell Leaper and Jonathan Gordon (Chappell et al. 1996) with funding from IFAW (International Fund for Animal Welfare). To distinguish porpoise clicks from the huge numbers of marine broadband clicks it compared, within a click, the energy at porpoise frequencies with the energy at lower frequencies. The IFAW detector also measured the amplitude of transient events, and detected clicks on both a bandwidth-related criterion, and an environmentally determined amplitude criterion.
The first POD did not embody any concept of a click as a spike in intensity. It continuously compared four bands of ultrasound and counted periods when the energy in the porpoise band exceeded each other band by separately user-defined ratios. It was, in effect, a simple analogue spectrum analyser. The spike concept was replaced with a single amplitude threshold to simplify the system avoid the complexities of a moving threshold based on non-porpoise spikes levels. It ran for 10 days, and stored, in different duration classes, counts of clicks that had met the user-defined selection criteria. Because of the flexible detection configuration a range of experiments on porpoise detection were possible.
Lessons learned from the Proto-POD:
Item 4: Dolphin clicks are not very distinctive except by their source level which is not available to a logger at an unknown distance from the source. Their detection could be approached by analysis of the characteristics of clicks in temporal clusters, but this appears to have serious limitations and temporal information that could be used to identify trains also appeared to be a valuable method for these animals.
Improving the Proto-POD
To extend the role of the POD to fishery interactions with dolphins and reduce (1) above alternative designs were considered.
A stereo system could show a common directional origin for groups of clicks and had been demonstrated by the IFAW team, now including Douglas Gillespie. This was rejected because of the impact on battery and memory size, and potentially poor performance on (1).
Analysis of the characteristics of groups of clicks (quasi-trains) appeared very powerful for porpoises, much weaker for dolphins and boat sonars, and costly on power, memory and complexity.
So a POD that logged click times, the T-POD, was designed, to make post-processing for trains possible. Various design compromises were required to increase running time on batteries, and were selected from the range of possible configurations:
A two filter analogue system was adopted, on the estimation that the known penalty of (2) could be outweighed by the second signal detection stage of train detection.
To optimally minimise the penalty of (2) flexibility of filter passband width ( 'Q') and centre frequency were included.
To enable dolphin detection without prior knowledge of the frequency of their clicks the system stepped through 6 user-defined sets of configuration each minute, each of which could have a different target and reference filter, ratio etc.
Click time and duration were the only parameters logged.
and evolution of versions
Successive versions of the T-POD had different features based on evidence from previous versions:
V1 T-PODs allowed testing of optimal Q values for target and reference filters.
In V2 and V3s:
In V4s and V5s:
Minimum threshold Sensitivity
All T-PODs had a user-controlled setting of this type with 16 settings, 0 being the lowest threshold. It was re-named sensitivity on a scale with 16 as the maximum sensitivity. This parameter allowed users to adjust PODs to match their own test results, if they fell within the available range. It also made possible settings that would reduce the number of 'dolphin-like' clicks. The measured threshold values were given in the POD specifications.
Ratio . click bandwidth
Detection occurs when the Target ( A ) band amplitude exceeds the Reference ( B ) band by this user-controlled ratio. Click bandwidth was = 7 ratio. So a narrowband click would exceed a high value of the ratio and be detected, and this setting was later described as a low click bandwidth. Simon et al, monitoring the Cardigan Bay MPA in Wales showed that a higher ratio (lower click bandwidth) was desirable in an area with many dolphins and porpoises as it improved the species discrimination, but it also involved a small loss of sensitivity, so user-control was retained for this selection criterion.
This feature was introduced in V4 and later PODs. If ON it had the effect of rapidly slightly raising the ratio criterion in response to high input levels from the filters. The half decay time of this adaptation, after the inputs fall below the threshold for this adaptation, was around 10ms. Its main effect was to reduce the number of clicks logged in the cluster of multipath replicates that follow a loud click, and to reduce the number of clicks logged during bursts of non-cetacean noise clicks. This markedly improved memory life and performance in noisy situations, but did not remove the need for users to identify periods in which noise may be impairing performance. It did not affect detection thresholds under 'normal' (quiet) conditions.
V4 and later T-PODs were standardised at manufacture. This adjusted the amplifier gain to give a 50% detection rate of a standard signal during rotation of the POD through 360 degrees.
During standardisation a variable level of background electromagnetic noise at POD frequencies (which are also radio frequencies) could produce significant changes in apparent sensitivity. This effect could be increased by the noise adaptation, so standardisation was carried out with noise adaptation OFF and the POD in a rotating, steel, screening housing. At the location used for standardisation the effect of varying levels of RF interference has been demonstrated repeatedly. The source is not generally known except that some active laptop computers do generate such interference.
The tank used to calibrate PODs was improved during the production of V4s. The amplitude of the loudest echo after the surface reflection was reduced to less than 10% of the direct path signal.
For C-PODs temperature compensation was introduced to the standardization after C-POD 400 as it became apparent that this was a significant factor.
End of the T-POD
The low-power Smart-Media memory cards used in V5 T-PODs were replaced by more power-hungry versions ( this change was never announced!) so that running times dropped substantially.
A larger battery pack and split housing were introduced and partly compensated for this. These memory cards then disappeared from the market and a memory adapter was added to allow a different card to be used.
A switch chip on the analogue board became hard to source. All available chips were bought by Chelonia and work was started on a V6 surface mount version of the analogue processor to allow more modern versions of this switch to be used.
The surface-mount V6 T-POD showed significant differences in performance from the V5 although the circuit models had not revealed this.
Meanwhile work was already under-way on a digital POD, the C-POD. At this time it also became clear that the T-POD digital processor would have to be changed as the newer memory card was now also only available in high-power versions (probably to meet ever-rising speed requirements in the camera market).
The workload involved in these issues and in developing and testing three boards for two different detectors, plus various revisions to the software that they would require, was going to create an excessively long gap before either could be supplied, so the incomplete V6 T-POD was abandoned.
The digital replacement the C-POD - addressed some major shortcomings in the type of detector used in T-PODs. These were:
The C-POD detector design was based on desk analysis trials on a large set of wideband (5 MHz sampling rate) recordings of dolphin clicks. The method selected is a wavelet that uses times and amplitudes of inflections and zero-crossings to produce a measure of the bandwidth of incoming sound in overlapping time windows of various lengths. Detection based on this 'micro-triggering' is better at finding relatively weak cetacean clicks in the context of louder ultrasonic noise with many amplitude spikes than 'macro-triggering' based on spectral analysis of longer time periods. Such spiky noise is commonly present in shallow seas.
The C-POD records, for every click, the dominant frequency of the first 10 cycles, the final zero-crossing interval, a bandwidth index, limited envelope information, the time of occurrence and the duration. For narrow-band signals this method allows accurate frequency read-out and very large dynamic ranges (the ratio of the weakest signal the system can handle correctly to the loudest).
This sounds very different from the T-POD, but is functionally surprisingly similar. Any tonal click that is louder than the atonal background noise will be recognised as tonal, and its frequency can be estimated. Signals weaker than the background will not be recognised. The background has the same function as the T-POD reference filter and the bandwidth measure has the same function as the ratio between filter inputs in the T-POD. In both instruments the electronic background within the instrument is most often louder than 'normal quiet' acoustic backgrounds. The acoustic background used within the C-POD is limited by a high-pass input filter, that can be controlled by the user, and by the inherent fall-off of transducer sensitivity at high frequencies, which cannot be controlled by the user, and by the sampling rate of the instrument.
Other changes included:
C-POD detection threshold
The adjustment of the detection threshold of the C-POD is done retrospectively using the amplitude of each click. Any value above 12 can be set correctly for any POD as all C-PODs can detect a signal that both meets the tonality (bandwidth) criterion and has an amplitude of 12 on the sound pressure level scale that records the maximum peak-to-peak amplitude within a click.
There is a reasonable case for logging only clicks louder than the uniform threshold value (12 on the C-POD SPL scale), but this has not been done because the 'personal best' i.e. weakest detectable clicks, of a set of C-PODs may show less variance than the sampling error in studies where detections are few. In such cases the statistical power of a study may be improved by using all detections even though the 'personal best' thresholds of different instruments are less uniform than the 'uniform threshold'.
The absolute SPL for scale values of 12 varies with frequency and can be viewed in Pascals in C-POD.exe by selecting this option on the view+ page of the menu. The frequency scale is given in the specification. (Note: The C-POD pressure scale runs from 3 to 255, with many loud clicks registering 255. Conversion of these point pressure values to decibels, a unit of relative intensity, not pressure, involves assumptions that are known to be invariably and substantially false in this situation, and the decibel itself is deprecated by the ICWM. )
To make this post-processing detection threshold uniform the amplitude scale of each C-POD is standardised as a temperature compensated average of the radial values obtained when a POD is rotated at 0.5rpm in a sound field with 300 clicks per second. The test signal is a square pulse of 12 cycles of a sine wave. The abrupt transition at each end is not ideal but has some practical advantages.
The train filter
The filtering effect of the train detection process is shown below:
Each vertical line represents the maximum peak-to-peak amplitude of a click in Pascals. All clicks in the raw data are shown in the lower panel. Only those clicks identified as belonging to a train are shown in the upper panel.
Trains are more or less regularly spaced series of similar elements. Cetacean trains show variation in the temporal spacing of clicks over time, and the similarity of the clicks is reduced by the changing orientation of the animal, propagation effects, and by changes in the click produced, especially in the case of broad-band dolphin clicks.
The train detection is based on a simple probability model of a train:
The probability, p of a click falling within some distance from the centre of the interval between the one before and the one after is determined by the Poisson distribution, the prevailing rate of arrival of clicks, a, the size of the interval, i, and the regularity of trains (that defines how close to the centre the click must be) so the probability of the whole identified train arising by chance from random sources will be the product of successive p values, and thresholds for acceptance of trains can be set using this.
There several important qualifications: a varies rapidly over time; the clicks in the train need to be subtracted from the count used to estimate a and this becomes a recursive process; low values of i are favoured; consequently estimation of the likelihood of false trains being found in real data sets is very unsatisfactory. Auto-correlation is a simple rigorous method for detection of trains with unvarying spacing of elements, and is of use in cetacean train detection within short time windows, but the search for a simple and rigorous method that can be applied to longer cetacean trains has been unsuccessful so far (although at a purely intuitive level it feels as though there should be one!), so the model, above, is in practice useful, but has to be empirically validated. These notes point to the shortcomings of train detection:
Black boxes v. transparency
Train detection and classification, in common with all but the simplest pattern recognition systems, is sufficiently complex that it is not possible to predict its performance from examination of the algorithm (Theodoridis and Koutroumbas, 2009) and because of this its development must require intensive testing and optimization using real data. By virtue of their complexity alone such processes are 'black boxes' that require external validation to determine their transfer function.
Manufacturers of electronic equipment in general either withhold or patent key design details of their instruments and even much simpler electronic instruments, from hydrophones to oscilloscopes, are also generally sold and used as black boxes with published empirical transfer functions. These are usually more accurate than any transfer function that might have been modeled using 'transparent' information on the components, circuits, logic etc. Even theoretically simple and transparent methods often produce unexpected errors in application so empirical validation is essential, and is, of course, also the basis of the scientific method.
Science has always made extensive use of non-transparent methods and progress has frequently depended on them e.g. litmus has been used in titrations to determine atomic weights even though no one knew how it worked or could predict the conditions in which it might fail. We proceed on the basis of intelligent scepticism.
Both train detection (that answers the question: Is there a sequence of clicks that came from a train source as opposed to arising by chance from independent sources?) and train classification (that seeks to identify the 'species' of source) for T-PODs went thorough a number of versions, each of which was fully retrospectively applicable to earlier data files. The C-POD train detection started with a provisional version (v1) that has been replaced by the KERNO classifier. This will not be improved as changes would require existing analyzed data sets to be re-analyzed.
Comparability of train detection versions
Early train classifiers used a set of train descriptors that were compared with values derived from known porpoise trains. They did not work well on some dolphin species that had less coherent trains, and the process was changed to one that aimed to exclude chance trains (wrongly extracted from noise) rather than identify trains matching porpoise trains.
The V1 C-POD train filter is based on the T-POD train filter and has a lot of room for improvement. A major part of that process is a shift to non-parametric methods that are more resistant to outliers which are usually non-cetacean clicks that were wrongly classified as falling within a train.
Improving Train classification encounter classifiers
Locations where classification errors are a problem are those where the number of false positives is comparable with or greater than the true positives this is seriously bad!
The rate of false positives is determined by the level of noise or interference. Interference is sound from sources that resemble the signal of interest e.g. dolphins can interfere with porpoise detection and vice versa.
To achieve lower mis-classification rates the data can be examined at a larger time-scale than is used by the KERNO classifier by extracting the properties of whole 'encounters' and classifying these. Encounters are defined as a sequence of minutes with a gap without trains of more than n minutes at each end. For different classifiers n and the quality score of qualifying trains can be defined.
The first such classifier developed is 'Hel1'. This is based on an analysis of data collected by the Hel Marine Station in Puck Bay. Visual examination of this data showed some false positives from:
All these features are readily incorporated into a classifier.
All classifiers present a trade-off between sensitivity and specificity. In this case the number of false positives from these sources is reduced by over 95% at the cost of a reduction in true positive clicks of about 13%. That's quite a good deal and makes it possible to use the data without visually checking all trains.
Some visual oversight is essential as we do not know all the possible sources of error and WUTS, for example, are still not well-described. However, visual validation is much quicker when false positives are few, because it is a reasonable assumption that they will occur in data sets that do not have true positives. ( if they only co-occurred with true-positives they would be much less of a problem! ) So files with very few detection can be selected for visual validation and examined very quickly.
Sediment transport noise
The first version of TPOD.exe required settings files to adapt the probability structure to data from deployments of PODs over sand which sometimes had very high rates of logging non-cetacean clicks. At the time we were not aware of the source of these clicks.
C-POD data from the same locations gives much more detail of these clicks and shows that they are often part of a population of clicks that shows trends in frequency. The C-POD can have much wider criteria for accepting clicks that the T-POD because it has a much larger memory and each click is logged with descriptors that allow, for example, porpoise-like clicks to be differentiated from non-porpoise-like clicks retrospectively. It gives a different kind of picture of ambient noise from that obtained by broad-band recording. In particular low frequency tones are excluded by higher frequencies as, for a tones with the same number of cycles, these are shorter.
The graphic above shows the distribution of frequency of tones logged in the Bristol Channel, a macrotidal estuary, over 30 hours. The frequency of tones is shown by colour (red = 20kHz, Violet = 140kHz), and a clear tidal pattern can be seen including even the difference between alternate tides.
We now realize that the source of most ultrasound logged in shallow water is sediment transport noise which has been identified and investigated by Thorne et al. (1986, 1988) who showed that it corresponds closely to rigid body radiation that arises when particles collide with each other, with fine sand producing tonal noise at porpoise frequencies. This has explained a lot of earlier T-POD data and clearly demonstrates that, with widespread tonal sources at porpoise frequencies:
Boat sonars, ADCPs
Boat sonars most often appear as clusters of tones that are very close to the frequency of the source, but sometimes quite strong harmonics are detected especially at the end of the multipath cluster. Because of their high source level and long duration, large clusters of tones, are commonly received from each sonar pulse. Embling has found that in 60m of water a C-POD could detect the sonar of a marine research vessel whenever it was within 1km.
Acoustic Doppler Current Profilers (ADCPs) operating at nominal frequencies far above the range of the POD sometimes also produce porpoise frequencies and can be wrongly identified as porpoises.
Work is at present under-way on a new C-POD that will have a different central processor (an anticipated low-power version of the current CPU has never been produced, so there is a risk that its production life may be suddenly ended if the manufacturer chooses to invest manufacturing resources in a more recent product). This instrument will also have:
Chappell, O., Leaper, R. and Gordon, J. (1996) "Development and performance of an automated harbour porpoise click detector." Rep. Int. Whal. Commn., 46, 587-594.
Simon, M., Nuuttila, H., Reyes-Zamudio, M. M., Ugarte, F., Verfuss, U., and Evans, P. G. H. (2010) "Passive acoustic monitoring of bottlenose dolphin and harbour porpoise, in Cardigan Bay, Wales, with implications for habitat use and partitioning." Journal of the Marine Biological Association of the United Kingdom doi:10.1017/S0025315409991226
Theodoridis, S., Koutroumbas, K. (2009) "Pattern Recognition" 4th edition. Academic Press.
Thorne, P.D., (1986) "Laboratory and marine measurements on the acoustic detection of sediment transport." J. Acoust. Soc. Am. 80 (3), 899-910.
Thorne, P.D. (1990) "Seabed generation of ambient noise." Journal of the Acoustical Society of America, 87, 149 153.