Stienessen et al.: Comparison of model types for prediction of seafloor trawlability in the Gulf of Alaska 189 
nonlinear relationships between the response and its pre- 
dictors and to optimize predictive performance (e.g., Elith 
et al., 2006, 2008). That is, the input data are weighted 
in subsequent trees rather than allowing each occurrence 
to have an equal probability of being selected. Boosted 
regression trees require the specification of 3 parameters: 
learning rate, tree complexity (the number of nodes in the 
tree), and bag fraction (Elith et al., 2006, 2008). We ran the 
BRT model, 
trawlability = S\, oblique + VRM + BPI, (3) 
following the guidelines discussed in Elith et al. (2008) 
by using the Bernoulli family for binary response data 
and reducing the learning rate to aim for over 1000 trees. 
For example, we used a learning rate of 0.001 shrinkage 
applied per tree, a tree complexity of 5, and bag fraction 
of 0.5. We assessed the value in simplifying the model by 
using the gbm.simplify procedure in the dismo package. 
Two-thirds of the data were selected to be part of the train- 
ing set, and the remaining one-third of the data were used 
as the test set (the same training and testing data sets 
used in the work with GLMs and GAMs). Model results 
were evaluated on the basis of values for D?, AUC, and the 
influence of the predictors. 
Random forests Seafloor trawlability was modeled with 
RFs by using the randomForest package, vers. 4.6-14 
(Liaw and Wiener, 2002), and caret package, vers. 6.0-84 
(Kuhn, 2019), in RStudio. In the RFs, random samples of 
predictor variables are used to generate many individual 
decision trees (forests) where each decision tree is con- 
structed by an individual bootstrap with replacement 
sample drawn from the original data. That is, the split 
determination at each node is based on the best-splitting 
variable from a randomly selected subset of variables. For 
classification, a weighted vote of the decision trees is used 
(Breiman, 2001). Data not in the bootstrap sample are 
out-of-bag data and are used as samples to be classified 
in each decision tree. Random forests require the spec- 
ification of 2 parameters: number of classification trees 
(number of bootstrap iterations) and the number of input 
variables to be used at each node. An RF has reliable 
predictive performance even when most predictive vari- 
ables are noisy; therefore, it does not require preselection 
of predictors (Diaz-Uriarte and Alvarez de Andres, 2006; 
Okun and Priisalu, 2007). 
On the basis of the results of a sensitivity analysis, we 
found that the caret package defaults (number of trees set 
at 500, and number of input variables set at 1) are appro- 
priate for our data and used them in the RF model: 
trawlability = S,, oblique + VRM + BPI. (4) 
Two-thirds of the data were selected to be part of the 
training set, and the remaining one-third of the data were 
used as the test set (the same training and testing data 
sets used in the work with the GLMs, GAMs, and BRTs). 
Model results were evaluated on the basis of values for the 
out-of-bag error estimate, AUC, and mean decrease in the 
accuracy of the predictors. 
Ensemble model An ensemble model was produced for 
estimation of seafloor trawlability by averaging predic- 
tions of the best GLM, GAM, BRT, and RF at each ras- 
ter cell (i.e., for each camera station). Two-thirds of the 
data were selected to be part of the training set, and the 
remaining one-third of the data were used as the test set 
(the same training and testing data sets used in the GLM, 
GAM, BRT, and RF) to compare the ensemble model fit to 
the individual models. 
Results 
A total of 47, 64, and 92 camera stations were sampled 
during the fine-scale surveys conducted with a Simrad 
ME70 during 2011, 2013, and 2015, respectively. Of these, 
only 38 (21 trawlable and 17 untrawlable) camera stations 
in 2011, 36 (22 trawlable and 14 untrawlable) camera sta- 
tions in 2013, and 56 (29 trawlable and 27 untrawlable) 
camera stations in 2015 produced data that could be used 
to estimate all 3 seafloor characteristics (Fig. 2). This was 
largely due to the fact that the video data were collected 
outside of the portion of the swath surveyed with a Simrad 
ME70 that was characterized by 35-50° incidence angles 
(i.e., there were no overlapping measures for Sj, oblique) 
at a number of camera stations. This effect, coupled with 
the unknown extent of trawlable and untrawlable areas 
within a grid cell a priori, produced an unbalanced final 
number of trawlable and untrawlable camera stations. 
The seafloor characteristic S,, oblique had a low correla- 
tion with the other 2 characteristics (VRM: coefficient of 
correlation [r]=0.15; BPI: r=0.05), and VRM and BPI also 
had a low correlation with one another (r=0.09). Values of 
S}, oblique associated with trawlable seafloor were signifi- 
cantly lower than those associated with untrawlable sea- 
floor when all 3 years of data were used (Fig. 3A). Values 
of VRM and BPI associated with trawlable seafloor were 
also lower than those associated with untrawlable seafloor 
when all 3 years of data were used, although the differ- 
ences were not significant (Fig. 3, B and C). 
Generalized linear model 
The best-fit GLM was a single-variable model in which 
S;, oblique is used. Only Sj, oblique was significant in 
the full model (P=0.0001). The variables VRM (P=0.11) 
and BPI (P=0.22) were not. Additionally, the AIC value 
for the single-variable model in which only Sj, oblique 
(AIC=151.5) is used was slightly less than that for the full 
model (AIC=153.4) (Table 1). 
Generalized additive model 
The best-fit GAM included a smoothed VRM term and a 
linearized S, oblique term. The BPI seafloor characteristic 
was not significant for the model. In the best model, the 
remaining terms were significant, linearized S, oblique 
was significant, AIC (147.5) was minimized, and D? (0.31) 
was relatively high (Table 2). Compared with this model, a 
