The effect of a multi-target protocol on cetacean detection and abundance estimation in aerial surveys

A double-platform protocol was implemented in the Bay of Biscay and English Channel during the SCANS-III survey (2016). Two observation platforms using different protocols were operating on board a single aircraft: the reference platform (Scans), targeting cetaceans, and the ‘Megafauna’ platform, recording all the marine fauna visible at the sea surface (jellyfish to seabirds). We tested for a potential bias in small cetacean detection and density estimation when recording all marine fauna. At a small temporal scale (30 s, roughly 1.5 km), our results provided overall similar perception probabilities for both platforms. Small cetacean perception was higher following the detection of another cetacean within the previous 30 s in both platforms. The only prior target that decreased small cetacean perception during the subsequent 30 s was seabirds, in the Megafauna platform. However, at a larger scale (study area), this small-scale perception bias had no effect on the density estimates, which were similar for the two protocols. As a result, there was no evidence of lower performance regarding small cetacean population monitoring for the multi-target protocol in our study area. Because our study area was characterized by moderate cetacean densities and small spatial overlap of cetaceans and seabirds, any extrapolation to other areas or time requires caution. Nonetheless, by permitting the collection of cost-effective quantitative data for marine fauna, anthropogenic activities and marine litter at the sea surface, the multi-target protocol is valuable for optimizing logistical and financial resources to efficiently monitor biodiversity and study community ecology.


Power Study Data Simulation
Based on a previous pilot study (miniSAMM), we can reasonably expect to obtain between 150 and 250 detections of Harbour Porpoises for the allocated effort for the "Double Plateform" experiment. We conducted data simulations for expected sample size between 150 and 250 by increment of 25. It can be anticipated that statistical power will depend on this sample size, with larger sample size leveraging more power to detect smaller bias.
Data simulation were carried out with a negative bias for the Megafauna protocol. The bias on detection could manifest itself in two different ways (see Figure 1).  Detections were modelled as Bernoulli trials: A difference between the two protocols can manifest itself in the intercept (β 0 ) or the slope (β 1 ). We considered only a negative bias for the megafauna protocol. In the simplest case, the probability to detect on the line (PerpDist = 0) was lower: where 0 < δ 0 < 1 is the bias.
We also considered a difference between slopes: where 0 < δ 1 < 1 is the bias. Since the sign of the slope is negative, this results in β mega Values for the bias parameters (δ 0 , δ 1 ) ranged between 0.05 and 0.50 by increment of 0.05.
For each combination of these parameter values, 500 datasets were simulated in R (R Development Core Team, 2015) version 3.2.3. These datasets were then analyzed with the package unmarked (Fiske & Chandler, 2011) to estimate the magnitude of a bias that was detectable with 80% statistical power at the 5% risk level.

Results
For all scenarios envisioned here, a bias in the slope parameter δ 1 was not detectable: in other words, even with a the sample size as large as 250 and a bias as large as 50%, statistical power was lower than 80% to detect a difference in slope. This result is unsurprising given the non-linear transform which maps the logit scale to the probability scale ( Figure 1) The magnitude of a detectable bias in the detection probability on the transect line for the megafauna platform depended on the detection probability on the transect line for the SCANS platform: the higher this latter probability, the smaller the detectable bias ( Figure 2). Unsurprisingly, the larger the sample size, the smaller the detectable bias. However, even in the most favorable scenario (sample size of 250 and 0.9 detection probability on the transect line for the SCANS platform), the smallest detectable bias was already larger than 15%. Thus, in this configuration, if the detection on the transect line for the megafauna platform is smaller than 0.9 × (1 − 0.15) ≈ 0.77, it would be detected at the 5% risk level with 80% power.
Values obtained from the pilot study, were logit −1 βscan 0 = 0.6 andβ scan 1 = −1.0. Thus, with a sample size of 150 detections, a bias of 35% was detectable with 80% statistical power. In other words, detection on the transect line for the megafaune platform must be smaller than 0.6 × (1 − 0.35) ≈ 0.4 to be statistically detectable.
However, a more surprising result from this pilot survey was that detection was actually better with the megafauna protocol ( Figure 3).

Observer Effects
Another important bias beyond that of the protocol is that of observers. We had a brief look at observer bias, again with the miniSAMM pilot study. Specifically, we fitted the following random effects model to the data: where (α scan , α mega ) i are observer effects. Observer effects were larger in magnitude than the estimated effect between the two protocols ( Figure 4). In fact, in the miniSAMM study, observers were nested within a protocol, and thus any protocol effect was confounded with observer effects. On Figure 4, the protocol effect is actually 0 while the previously identified difference (Figure 3) is visible as an observer effect. While the miniSAMM design prevents the joint estimation of an observer and protocol effects, the miniSAMM data nevertheless suggest that any difference between the two protocols may be dwarfed by between-observers variability.