Campbell et al.: Attraction and avoidance behaviors of fish in response to proximity of underwater vehicles 227 
abundance increased with increasing number of tran- 
sects. Because fish of the local reactive guild were not 
attracted to the vehicles but abundance increased as a 
function of vehicle range and over transects, vehicle pas- 
sage could have altered the behavior of the fish, causing a 
coincident increase in detectability by the MOUSS plat- 
forms. In other words, detection rather than abundance is 
potentially changing, at least from the perspective of the 
MOUSS cameras. 
Detection probability and sighting functions (i.e., range 
to positive identification of fish) have been shown to be 
critical for estimation of fish densities for strip-transect 
and distance sampling methods (Sale and Sharp, 1983; 
Ensign et al., 1995; Cheal and Thompson, 1997), even 
for species that have minimal or no avoidance behaviors 
(Kulbicki and Sarramégna, 1999). Although the functions 
developed in this work add critical information about how 
relative fish abundance changed prior to vehicle arrival, 
vehicle-specific sighting functions are equally important 
to develop even for species exhibiting minimal or no behav- 
ioral response. Further, it is important to understand that 
the MOUSS platforms in our experiment have their own 
detection probability and sighting functions; therefore, 
what is presented here is relative to the MOUSS platform 
observations. Thus, our modeling does not necessarily lead 
to unbiased estimates of abundance or biomass. Further 
research that could establish fish density in a known area, 
and therefore establish the standard in which the optical 
systems were deployed, would be a useful but extremely 
difficult undertaking. Finally, all of the key components 
necessary to estimate fish density will likely have import- 
ant covariates to consider, variables such as light, water 
clarity, and habitat complexity. Light and water clarity 
were fairly constant over the course of the experiment; 
however, habitat complexity varied among sites. Both of 
these findings may or may not occur in basin-wide surveys 
but are critical considerations for those types of surveys 
(e.g., in the Gulf of Mexico). 
Although we were able to estimate functions that were 
based on vehicle range and that included important covari- 
ates to help evaluate fish response (e.g., habitat complexity), 
we were unable to connect those responses to any specific 
vehicle stimuli (e.g., noise or light). Sensory mechanisms of 
fish that detect the sound, light, and motion of a vehicle are 
all critical components that likely incite fish reaction and 
potentially bias observations in undetermined directions 
and magnitude. For instance, in an experiment designed 
to observe soniferous fish behavior, ROV noise induced a 
strong negative behavioral response that affected observa- 
tions during sampling (Rountree and Juanes, 2010). Simi- 
larly, responses to artificial lighting range from attraction 
to avoidance and are species specific (McIninch and Hocutt, 
1987; Marchesan et al., 2005; Raymond and Widder, 2007). 
Furthermore, the order of stimuli detection and their asso- 
ciated thresholds in fish are uncertain. For instance, the 
sound produced by motors and propellers or the vibration 
of tow cables generate noise that could stimulate the lat- 
eral movers guild fish and cause a response prior to the 
fish detecting vehicle lights. Similarly, positioning beacons 
generate sound but generally operate at higher frequencies 
than fish can detect (Mann et al., 2001; Popper, 2003; Stoner 
et al., 2008). 
Ultimately, noise-inducing equipment, such as tethers, 
tow cables, and propellers are critical components of vehi- 
cles; therefore, experiments should include a capacity to 
measure relevant stimuli in order to relate them to fish 
response. For instance, future iterations of this type of 
experiment and large-scale ocean observation systems 
(Rountree et al., 2020) would benefit from having a suite of 
acoustic sensors and light meters to identify specific vehi- 
cle noise and light production signatures. Light meters 
and turbidity sensors for detecting ambient conditions in 
the environment would also be useful for estimation of 
vehicle- and condition-specific sighting functions. Finally, 
experimental methods that place sampling platforms 
at specific locations and on specific habitat would likely 
result in more interactions with target species and enable 
development of species-specific range functions. Cameras 
that have increased FOV (e.g., 360° or full spherical) 
might be useful, given that it has been reported that they 
increase fish detection (Kilfoil et al., 2017; Campbell et al., 
2018). Additionally, full-spherical cameras would be useful 
in vehicle detection when they transit behind or above the 
intended transect line, and thus more gear interactions 
could be captured. 
The original statistical analysis was envisioned to cre- 
ate single-species models that include vehicle type as a 
variable in order to facilitate direct comparisons of vehi- 
cle performance. Logistics of the ship, however, caused us 
to stage the experiment in 2 separate periods. Therefore, 
in the experiment, each vehicle was tested in somewhat 
different habitats that had varied fish assemblages and 
densities, making it impossible to expose each vehicle to a 
standardized set of conditions. Because of the difficulty of 
sampling different sites, we created the habitat complex- 
ity variable in an effort to standardize site conditions and 
clarify changes in abundance that were associated with 
vehicle passage. Additionally, species-specific interactions 
in the coincident sampling volume in space and time were 
rare, thus making single-species models problematic to 
estimate. Other at-sea conditions also created situations 
where the AUV in particular was not deployed as fre- 
quently as the other vehicles and therefore had fewer 
interactions with target species. For these reasons, we felt 
that a direct comparison that used a singular model (i.e., 
inclusion of a vehicle variable) was not defensible. Impor- 
tantly, our observations were that each vehicle had good 
and bad traits for sampling different habitats and had 
differential utility depending on target species. Thus, we 
made an effort to not compare vehicles qualitatively (i.e., 
good or bad). We envision that the outcomes of our work 
potentially would help researchers select a vehicle that 
best fits the target environment and species they intend 
to sample. 
Critically, underlying fish density could not be con- 
trolled; therefore, statistical analysis, model estimation, 
and resultant explanatory capacity were shaped around 
the data that we were able to obtain. Recent efforts to 
