270 
Fishery Bulletin 114(3) 
NYB4 
NYB5 
NYB8 
Digitally sized scallops (AUV survey) 
Manually sized scallops (dredge survey) 
NYB6 
NYB9 
Shell height (mm) 
NYB7 
NYB10 
Figure 6 
Histographs of shell height distrihution for digitally sized sea scallops (Placopecten magellanicus) 
photographed with an underwater camera of an autonomous underwater vehicle (AUV) and for sea 
scallops collected with a commercial dredge in fishing operations undertaken at the same time and 
at the same locations in the New York Bight area. 
tors could complete the observation of 203,000 images in 
98 man-hours (a rate of 2,000 images per hour). Walker 
(2013) showed that the hours required to complete an- 
notation of a set of images were directly related to the 
number of scallops and associated fauna. 
Over 200,000 bottom photographs were obtained in 
this study, and all were examined by annotators trained 
to count scallops and measure their shell height. We 
found this manual step to require many hours of effort 
and some expense. We investigated statistical approaches 
by repeated random sampling simulations and text book 
formulae (Zar 1999, p. 109) that can be used to gauge the 
loss in precision by examining only a random subset of 
images. Overall, we found the mean density to be 0.075 
scallops/image, with a standard deviation of 0.35. These 
values were sufficient to compute the standard error of 
the mean density from a smaller random subset of im- 
ages with the formula for the standard error, SE = s/Vn, 
where s is the standard deviation and n is the number of 
photos. For example, a random subset of 40,000 images 
would have a standard error of 0.35/^40,000=0.00175, 
giving a 95% confidence interval (i.e., twice the standard 
error) of ±0.0035. This bound is at ±5% of the mean value 
(i.e., the relative error) obtained from all photographs 
and can be the expected precision when examining a ran- 
dom sample of only 20% of the images that we collected. 
Sampling half that number (20,000) increases the rela- 
tive error to 6.6%, while doubling it to 80,000 decreases 
the relative error to 3.3%. Because imagery-based as- 
sessments typically generate large numbers of images. 
