190 Fishery Bulletin 119(2-3) 
175°W 170°W 165°W 160°W 155°W 150°W 145°W 140°W 
Alaska 
Unimak Island 
165°W 160°W 
155°W 
Gulf of Alaska 
Kodiak Island 
Survey year 
» 2011 
© 2013 
yx 2015 
150°W 
Figure 2 
Map of the locations of camera stations where fine-scale multibeam surveys were conducted in the 
Gulf of Alaska during the summers of 2011 (circles), 2013 (diamonds), and 2015 (stars). The col- 
lected multibeam sonar data were used to derive seafloor characteristics used in models to predict 
trawlability of habitat utilized by rockfishes (Sebastes spp.). 
model that included a smooth function for Sj, oblique had 
a higher AIC (150.5), although the D? (0.31) was the same. 
A model that included all 3 seafloor characteristics and 
assigned a smooth function to them had an even higher 
AIC (154.4), and the D? (0.25) was lower. 
Boosted regression tree 
Results from the use of the gbm.simplify procedure 
indicate that the full BRT model (with all 3 variables) 
could be simplified to a 2-variable model by dropping 
VRM. This simplification resulted in a more parsimoni- 
ous model without degradation of model fit. That is, the 
removal of VRM did not adversely affect model predictive 
performance. Deviance explained (0.31) was the same for 
the 2 models, and test AUC increased by 0.01 (to 0.65) in 
the simplified model. The predictor Sj, oblique influenced 
both models the most (68% in the simplified model and 
59% in the full model), followed by BPI (simplified: 32%; 
full: 28%). The VRM influenced the full model by only 
13% (Table 3). 
Random forest 
The global accuracy (proportion of correct classification of 
trawlability) for this model was 0.655. Restated, the RF 
incorrectly classified trawlability 34.5% of the time (out- 
of-bag error estimate=0.345). The seafloor characteristic 
Sj}, oblique contributed the most to the model’s predic- 
tion performance (mean decrease accuracy=0.180), but 
permuting VRM (mean decrease accuracy=0.013) and 
BPI (mean decrease accuracy=0.006) did not obstruct 
the model substantially. That is, VRM and BPI were not 
helpful predictors. When either VRM or BPI was per- 
muted in a way that allowed its distribution to remain 
the same but assigned its specific observations randomly 
to the data, there was minimal loss of accuracy in 
classification. 
Model evaluation 
Results from all full models indicate a strong relationship 
between seafloor trawlability and S,, oblique. Additionally, 
S;, oblique was the most important predictor in all models. 
The seafloor characteristic VRM was significant in the 
GAM, and BPI contributed to the performance of the BRT. 
The GLM and GAM produced smooth or linear responses, 
but the significance of VRM was not consistent in these 2 
models (Fig. 4). The response curves for the BRT and RF 
were similar to each other, indicating that the use of these 
2 models revealed the same relationships in the data 
(Fig. 4), although the different learning methods between 
the 2 models resulted in different associations. In the RF, 
the initial variable is selected at random, and over many 
iterations the choice typically becomes obvious (e.g., Sj, 
oblique). However, when similar predictions result from 
