Stienessen et al.: Comparison of model types for prediction of seafloor trawlability in the Gulf of Alaska 193 
Trawlability 
T 
-25 -5 -0.02 
S,, oblique (dB) 
| 
0.28 
VRM 
Figure 4 
Relationships between explanatory variables of models—oblique incidence backscatter (S), 
oblique), vector ruggedness measure (VRM), and bathymetric position index (BPI)—and trawlabil- 
ity of habitat utilized by rockfishes (Sebastes spp.). The generalized linear model (GLM) depicts 
the single-variable model in which S,, oblique was used, the generalized additive model (GAM) 
includes a smoothed VRM term and a linearized Sj, oblique term, the boosted regression tree 
(BRT) is the full model, and the random forest (RF) is the full model. Response curves that are 
horizontal lines for the entire extent of variables in the GLM and GAM models indicate that these 
variables were not significant in the model. Gray dashed lines indicate 95% confidence intervals. 
Data used in models were collected during fine-scale multibeam surveys in 2011, 2013, and 2015 
in the Gulf of Alaska. 
study, characteristics derived from multibeam sonar data 
were used to predict seafloor trawlability for both locations 
of camera stations and for historic locations of trawl tows 
of the AFSC bottom-trawl survey. The latter was defined 
as areas where a bottom-trawl survey had previously been 
conducted, either successfully with little or no gear damage 
(trawlable) or unsuccessfully with extensive gear dam- 
age (untrawlable). For both camera station and trawl tow 
locations, Pirtle et al. (2015) found that a model in which 
only S;, oblique is used to be among the most discriminat- 
ing single-variable models for predicting trawlability. Fur- 
thermore, the addition of other variables did not improve 
predictive discrimination for data from tow locations. 
The relevance of VRM and BPI in predicting trawlabil- 
ity is less clear. Results of our study indicate that the 
addition of VRM and BPI does not improve on the effec- 
tiveness of a single-variable linear model in which only Sj, 
oblique is used to predict trawlability. Similarly, in other 
studies, seafloor rugosity was a poor discriminator 
between trawlable and untrawlable seafloor at Snake- 
head Bank (Weber et al., 2013), and neither VRM nor BPI 
improved on the performance of the single-variable linear 
model in predicting trawlability at tow locations over 
larger portions of the GOA (Pirtle et al., 2015). However, 
the addition of both VRM and BPI improved the effective- 
ness of the single-variable linear model in predicting 
trawlability at camera stations in the study of Pirtle et al. 
(2015). They found that including S;, oblique, VRM, and 
BPI resulted in 1 of the 2 most discriminating GLMs for 
explaining trawlability in the GOA (AIC=32.1, D?=0.55). 
They calculated VRM slightly differently than we did, but 
that difference does not appear to explain the discrepancy 
in results. When we updated the data from 2011 used in 
the Pirtle et al. (2015) study with VRM calculated with 
consistent units (i.e., updated with positions in meters of 
x and y coordinates in the Universal Transverse Mercator 
system and depth in centimeters), the relationship 
between trawlable and untrawlable areas determined by 
using VRM values did not change. The VRM values for 
trawlable areas were still significantly lower than values 
for untrawlable areas (Fig. 3D), although the performance 
of the model decreased slightly (AIC=33.4, D?=0.45). 
