188 
Fishery Bulletin 119(2-3) 
analysis window (BPI) was outside of the 20 m. Single S;, 
oblique, VRM, and BPI values were derived for each cam- 
era station by taking an average of each value collected 
along the camera path. 
Video data Seafloor substrate was classified as either 
trawlable or untrawlable during review of video images 
by an experienced AFSC analyst. Designation of traw- 
lability was based on whether the standard 4-seam 
Poly-Nor’Eastern bottom trawl used by AFSC in biennial 
bottom-trawl surveys (Stauffer, 2004) could successfully 
trawl in a given area. Specifically, analysts defined untraw- 
lable areas in video images as any substrate containing 
boulders higher than 20 cm off bottom, an elevation that 
corresponds to the height of the roller gear on the footrope 
of the Poly-Nor’Eastern trawl (Rooper et al., 2012). They 
also classified areas in video images as untrawlable if the 
seafloor contained bedrock with vertical relief or rugged- 
ness greater than 20 cm that would likely prevent the bot- 
tom trawl from passing over it without damage to the net 
from seafloor contact. 
Modeling seafloor trawlability 
Ultimately, each camera station was classified as either 
trawlable or untrawlable and was characterized by S), 
oblique, VRM, and BPI. The analyses done with models 
included only the stations for which data for all 3 variables 
were produced. Occasionally it was not possible to calcu- 
late S;, oblique at a camera station. Although the camera 
deployment paths fell within the area covered during a fine- 
scale multibeam survey (Fig. 1), the positions of the camera 
deployments did not always overlap with the portion of the 
survey swath for which a Simrad ME70 collected data at 
35-50° incidence angles. The effectiveness of using the 
seafloor characteristics derived from multibeam sonar data 
to discriminate between trawlable and untrawlable habi- 
tat was tested with 4 different kinds of predictive models: 
GLMs, GAMs, BRTs, and RF's. The GLMs were used to build 
upon, and allow direct comparisons to, the work of Pirtle 
et al. (2015). The GAMs are similar to the GLMs but allow 
any nonlinear relationships between the seafloor charac- 
teristics to be expressed. The BRTs and RFs have machine 
learning strength for classification and allow inclusion of 
nonlinear components. 
Generalized linear models Seafloor trawlability was mod- 
eled with logistic regression by using the GLM: 
trawlability = By + (By, x Sj, oblique) + (Bz x VRM) 
+ (B, x BPI) + E, (1) 
where By ,23 =the parameters estimated by the model, 
and 
E = error; 
in the GLM, the logit link function p(Y)=e(Y)/1+e(Y) is 
used for binary response data (McCullagh and Nelder, 
1983). The logit link function maps probability values (p) 
between 0 and 1 for any given real number Y between 
negative and positive infinity. Seafloor characteristics 
were standardized prior to analysis by subtracting the 
mean and dividing by the standard deviation. 
We ran a GLM with all 3 years (2011, 2013, and 2015) 
of data. Backward variable selection was then performed 
on the data from all 3 years by fitting the model and drop- 
ping terms on the basis of an insignificant P-value for the 
model term (i.e., P=>0.05) and the Akaike information crite- 
rion (AIC). The predicted responses of models with AIC val- 
ues within 2 digits are not considered different (Burnham 
and Anderson, 2002). Deviance explained (D?), an analy- 
sis of deviance test comparing the model results to a null 
model (level of significance=0.05), and the area under the 
receiver operating characteristic curve (AUC) were also cal- 
culated for comparison with the 3 other types of predictive 
models. In general, an AUC value of 0.5 indicates that the 
model cannot discriminate between trawlable and untraw- 
lable seafloor; AUC values of 0.7—0.8 are acceptable, values 
of 0.8—0.9 are good, and values >0.9 are excellent (Hosmer 
and Lemeshow, 2000). To evaluate the performance of the 
GLMs, we fit a GLM to two-thirds of the data (randomly 
drawn without replacement) and used the remaining one- 
third of the data for testing. Closer agreement between the 
area under the receiver operating characteristic curve for 
the test data set containing out-of-sample data (test AUC) 
and the area under the receiver operating curve for the 
training data set (training AUC) indicates better predic- 
tive performance by the model. 
Generalized additive models Seafloor trawlability was also 
modeled with GAMs by using the mgcv package, vers. 1.8-28 
(Wood, 2017), in RStudio, vers. 1.2.5019-6 (RStudio, Boston, 
MA). These models assume that the effects of trawlabil- 
ity are additive and can use smooth functions (penalized 
regression splines) to model each predictor on the response 
variable (Wood and Augustin, 2002). Best-fitting models 
were determined in a backward stepwise fashion beginning 
with the GAM model: 
trawlability = s(S), oblique) + s\VRM) + s(BPI)+£E, (2) 
where the function s(x) indicates a smooth effect of each 
predictor. In the GAM, the logit link function is used for 
binary response data. Terms were dropped on the basis of 
a nonsignificant test for the F-term in the nonparamet- 
ric effects, and terms were made into a linear function 
instead of a smooth function on the basis of a signifi- 
cant test for the F-term in the parametric effects. Model 
results were compared on the basis of values for DACs 
and training AUC. To evaluate the performance of the 
GAMs, we fit a GAM to two-thirds of the data and used 
the remaining one-third of the data for testing (the same 
training and testing data sets used in the work with the 
GLMs). 
Boosted regression trees Seafloor trawlability was mod- 
eled with BRTs by using the dismo package, vers. 1.1-4 
(Hijmans et al., 2017), in RStudio. Unlike traditional 
regression tree methods, BRT boosting is used to combine 
large numbers of simple tree models adaptively to display 
