194 
Fishery Bulletin 119(2-3) 
When 2 more years of data (i.e., data from 2013 and 2015) 
were added to the model, performance decreased further 
(AIC=153.4; D?=0.19; Table 1). 
Pirtle et al. (2015) postulated that VRM and BPI were 
important in predicting trawlability at locations of cam- 
era stations but not at trawl tow locations because at 
camera stations the range of seafloor types that was sam- 
pled (i.e., both trawlable and untrawlable) was broader 
than the range sampled at tow locations where surveys 
were designed to target trawlable seafloor. However, with 
the addition of the data from 2013 and 2015 to the lin- 
ear model that incorporates a broader range of seafloor 
types, VRM and BPI were no longer necessary predictors. 
The portions of the GOA where data were collected from 
camera stations were larger in 2013 and 2015 than in 
2011 (Fig. 1). That is, they were collected farther west 
(i.e., west of Umnak Island: 52°38’N, 169°14’W) and far- 
ther east (i.e., east of Kodiak Island) (Fig. 2). This differ- 
ence indicates that VRM and BPI are not as effective in 
GLM predictive discrimination of trawlability beyond the 
geographic area sampled in 2011. 
The power of the use of seafloor characteristics derived 
from multibeam sonar data to predict seafloor structures 
is dependent on spatial scale (Wilson et al., 2007). That 
is, the predictive power of the seafloor characteristics is 
dependent on the areal extent used to initially create the 
model. The predictive power of the characteristics is also 
S, oblique (dB) 
2011 
more pronounced when they have a greater range of val- 
ues. The ranges of BPI values extracted in 2013 and 2015 
were not as large as the range of those extracted in 2011 
(Fig. 5). The VRM and BPI may be more useful as predic- 
tors when more “extreme” bottom types are encountered, 
and they may not have as much predictive power when 
their extracted values are closer to the overall mean. The 
areal extent of the study site used by Pirtle et al. (2015) 
to create the model likely affects the applicability of the 
model over a larger extent of the GOA. 
Even though VRM and BPI can be omitted from the 
GLMs and RFs without much loss in effectiveness at 
defining trawlable versus untrawlable seafloor in the GOA, 
they were independently useful for predictive discrimina- 
tion of trawlability in the GAMs and BRTs. That is, VRM 
was included in the most parsimonious GAM, and BPI was 
included in the most parsimonious (i.e., simplified) BRT. 
It is unclear why predictive performance with VRM and 
BPI is not consistent among the GAMs and BRTs, because 
these characteristics are not strongly correlated. They 
measure different seafloor properties; therefore, there 
should be no confusion about which characteristic should 
be dropped from the model. The addition of either VRM or 
BPI to the respective nonlinear models results in reason- 
able test AUC scores for the out-of-sample sites. 
Our results indicate that all the models identify S, 
oblique as the primary driver for predictive discrimination 
2013 2015 
Figure 5 
Box plots for the seafloor characteristics, (A) oblique incidence backscatter (Sj, oblique), (B) vector rugged- 
ness measure (VRM), and (C) bathymetric position index (BPI), derived from data collected with a Simrad 
ME70 multibeam echo sounder in 2011, 2013, and 2015 in the Gulf of Alaska. Seafloor characteristics 
were used to predict trawlability of habitat utilized by rockfishes (Sebastes spp.). The boxes show the 
first quartile (lower line), median (middle line), and third quartile (upper line). The whiskers extend to 
the minimum and maximum values. Gray circles represent outliers, and the x in the middle of each box 
indicates the mean. 
