Comment on “Dolphin and porpoise detections by the F-POD are not independent: Implications for sympatric species monitoring” [JASA Express Lett. 4, 031202 (2024)]

The key acoustic issue is that dolphins, boat sonars, and sediment in suspension can and do make clicks with acoustic properties that fall well within the much narrower range of porpoise click properties. So any acoustic classifier of clicks, whether an F-POD or a conventional broadband recorder, that simply seeks narrow-band high frequency, porpoise-type, clicks may give false positives when those sources are present.

The KERNO-F classifier, used to analyze F-POD data, and studied in this paper, aims to keep the risk of false positives very low, even at the cost of losing some sensitivity, because, for example, false positive porpoise detections from dolphins (or other sources) could completely mask a catastrophic decline in a rare porpoise.

In some locations, for example, some rivers, porpoises are known to be absent and the KERNO-F classifier can be given this as prior information, and it can then relax the criteria that were otherwise used to minimize the risk of false classification of porpoises clicks as dolphin clicks. Advisory text in red is shown at the top of the software screen if this “No porpoises” option is visited, stating “This increases the sensitivity of dolphin detection slightly … provided porpoises are really absent!”

These options (no dolphins, no narrow-band high frequency species, no sonars) have been treated in this paper as creating independent classifiers, but actually the only thing they do is force all clicks that could have belonged to more than one category into the remaining category.

By contrast, an “independent classifier” would have to be constructed to reject nearly everything that could possibly be in either category or it would risk very high false positive rates in adverse environments. In doing this, it would not be independent, but would be working against some fixed, and usually unstated or unknown, assumption of some level of prevalence of the alternative categories, and that would also bias the characteristics of what it did accept.

The authors note that, in F-POD data from Ukraine, the “KERNO-combined” and “KERNO-separated” classifications resulted in a single dolphin DPM becoming almost 700 dolphin DPMs, respectively [Fig. 1 from Cosentino (2024)]. However, if validation of those detections had been carried out, it would have shown that they were all porpoises that had been wrongly forced into the dolphin category by making a prior assertion that they must belong there, and their inclusion in Fig. 1 actually illustrates how misleading this approach is.

The anomalous results they find in the file from Turkey are also completely predictable; the raw data had 200 million clicks. The KERNO-F classifier found 12 k sonar DPM but only 31 porpoise DPM. In that site, there is clearly a high risk of porpoise-like clicks from sonars being misclassified as porpoises. If the authors had asserted “No sonars,” close to 1 million porpoise clicks in 6 k porpoise DPMs would have been found. These would, predictably, be nearly all false, but would have given even more extreme values in Fig. 1.

In essence the idea of using independent classifiers only makes sense when each species class has a universally present unique feature, which would allow the “classifier” to become a “counter” of cases with those features. This is not plausible for any of the classes here. The problem of interactions is actually more extensive than raised in Cosentino (2024) as there are also interactions with sediment transport noise and boat sonars. We take the view that:

1. In order to keep false positives low, classifiers of overlapping categories, such as dolphins, porpoises, sonars, etc., should NOT be independent. This is true of broadband loggers as well as PODs, and a published example of how it goes wrong is referenced by Ivanchikova and Tregenza (2023) where an independent click-based spectral classifier probably delivers many false porpoise detections from sediment transport noise (but source data were not made available).

2. The interactions between these identifiable sources are actually quantifiable, and Ivanchikova and Tregenza present the first quantitative estimate of two of those interactions. Further work may show that some useful scaling factors can be established in that way and can be used predictively.

3. Until then, determining how often interacting categories coexist is possible, but it does require experienced human validation of the data, including all those click trains placed in the “unclassified” category, as such analysis can out-perform the currently available classifiers. Details of this are given in Ivanchikova and Tregenza, and such an approach is needed for studies of interactions between dolphins and porpoises, between boat sonars or pingers and cetaceans, and between some types of ambient noise, particularly sediment transport noise, and cetaceans.

4. A rough assessment of the risk of substantial bias is possible: (a) Assume the temporal distribution of the species is independent. (b) Find the fraction of minutes with a species group. (c) Look at the whole file to see if the two species are actually clustered together in time as this would increase the interaction and bias. (d) If they are not clustered, then consider the fraction as an estimate of the fraction of minutes with a bias against identification of the presence of another species. (e) If the species are clustered together in part of the overall time, then the fraction of minutes should be taken from that part of the file duration.

5. The supporting information in Ivanchikova and Tregenza already states: “potential risk factors can be usefully predicted from the nature of the detection process …” and includes all the issues identified in this paper and more. Chelonia provides information in explanatory videos on when and how such analysis should be carried out and will respond to this paper by adding further resources for users of these instruments.

N.T. is a manufacturer of the F-POD devices. J.I. has no conflicts of interest to disclose.

The minimal dataset, which is needed to reach the conclusions drawn in the manuscript, is located in the Figshare Repository via https://doi.org/10.25452/figshare.plus.23982288.