192 
Fishery Bulletin 119(2-3) 
Ensemble model 
The ensemble model was produced by averaging the pre- 
dictions of the best models (i.e., the single-variable GLM 
in which S;, oblique was used, the GAM that included a 
smoothed VRM term and a linearized Sj, oblique term, the 
simplified BRT, and the full RF) for each camera station. 
The fit for the ensemble model indicates that the model 
was reasonably effective in separating trawlable and 
untrawlable habitat (test AUC=0.70). The ensemble mod- 
el’s performance in predicting trawlability with the test 
data was better than the performance of the simple BRT 
(AUC=0.65; Table 3) and RF (AUC=0.63), it was equal to 
Table 2 
Comparison of results from generalized additive models 
used to predict seafloor trawlability of habitat utilized 
by rockfishes (Sebastes spp.) at locations of camera sta- 
tions surveyed during 2011, 2013, and 2015 in the Gulf 
of Alaska. In the models, s(x) indicates a smooth effect of 
the predictor. Models are listed in order of decreasing accu- 
racy of the predictions of trawlability. Predictor variables 
are seafloor characteristics derived from multibeam sonar 
data: oblique incidence backscatter strength (S,, oblique), 
vector ruggedness measure (VRM), and bathymetric posi- 
tion index (BPI). Values of Akaike information criterion 
(AIC), deviance explained (D), area under the receiver 
operating curve for the training data set (training AUC), 
and area under the receiver operating curve for the test 
data set containing out-of-sample data (test AUC) are pro- 
vided for each model. 
Training Test 
Model AUC AUC 
S, oblique + s(VRM) 0.86 0.70 
s(S;, oblique) + s(VRM) 0.87 0.70 
s(S;, oblique) + s(VRM) 0.88 0.68 
+ s(BPI) 
the performance of the best GAM (AUC=0.70; Table 2), 
and it was slightly lower than the performance of the best 
GLM (AUC=0.75; Table 1). 
Discussion 
All 4 types of predictive models produced similar overall 
results: Sj}, oblique was the most significant predictor to 
discriminate between seafloor designated as trawlable 
or untrawlable through analysis of video images, and all 
tested models were moderately effective in predicting 
trawlability at camera stations across the shelf in the 
GOA (D? range from 0.17 to 0.31, and test AUC values 
range from 0.63 to 0.73). Results from the GAM and BRT 
indicate that one predictor can be omitted, but they dis- 
agree as to whether the predictor should be BPI (GAM) or 
VRM (BRT). Results from the GLM and RF indicate that 
both BPI and VRM could be omitted without much loss 
in the accuracy of the classification of trawlable versus 
untrawlable habitat. 
The prominence of Sj, oblique in predicting trawlability 
corroborates the results of previous work (Jagielo et al., 
2003), specifically those studies on trawlability in the 
GOA (Weber et al., 2013; Pirtle et al., 2015), but over a 
much larger geographic extent. Additionally, the lower 
values of S;, oblique associated with trawlable areas com- 
pared with those associated with untrawlable areas indi- 
cate a lack of strong scatterers, like hard rock and boulders, 
in trawlable habitat (Jackson and Richardson, 2007; 
Lamarche et al., 2011). Weber et al. (2013) found that 
trawlable areas had lower S,, oblique values than untraw- 
lable areas at Snakehead Bank in the GOA. Furthermore, 
of the measured seafloor characteristics, Sj, oblique had 
the best performance in predicting trawlability at the 
Snakehead Bank study site (Weber et al., 2013). Pirtle 
et al. (2015) also found that untrawlable areas had higher 
SS}, oblique values than trawlable areas at many locations 
between Akutan Island (53°43’N, 164°58’W) and the east- 
ern side of Kodiak Island (57°18’W, 151°30’N). In their 
Table 3 
Comparison of results from boosted regression trees, with 2 or 3 variables, used to predict sea- 
floor trawlability of habitat utilized by rockfishes (Sebastes spp.) at locations of camera stations 
surveyed during 2011, 2013, and 2015 in the Gulf of Alaska. Predictor variables are seafloor 
characteristics derived from multibeam sonar data: oblique incidence backscatter strength 
(S;, oblique), vector ruggedness measure (VRM), and bathymetric position index (BPI). Values 
of deviance explained (D”), area under the receiver operating curve for the training data set 
(training AUC), and area under the receiver operating curve for the test data set containing out- 
of-sample data (test AUC) are provided for each model. The dash in the table signifies that the 
corresponding predictor variable (VRM) was not used in the associated model. 
S;, oblique VRM BPI Training Test 
Model D influence influence influence AUC AUC 
S,, oblique + BPI 0.31 68.2 31.8 0.86 0.65 
S;, oblique + VRM + BPI 0.31 59.3 27.6 0.88 0.64 
