Authors Lisa Wedding Mary M. Yoklavich
License CC-BY
Vol. 540: 235–250, 2015 MARINE ECOLOGY PROGRESS SERIES
Published November 26
doi: 10.3354/meps11442 Mar Ecol Prog Ser
OPEN
ACCESS
Habitat-based predictive mapping of rockfish density
and biomass off the central California coast
Lisa Wedding1, 2,*, Mary M. Yoklavich1
1
Fisheries Ecology Division, Southwest Fisheries Science Center, National Oceanic and Atmospheric Administration,
110 Shaffer Rd., Santa Cruz, California 95060, USA
2
Center for Ocean Solutions, Stanford University, 99 Pacific Street, Suite 555E, Monterey, California 93940, USA
ABSTRACT: Understanding the association between components of habitat and fish distribution
and abundance is important in order to achieve accurate stock assessments. We developed gen-
eralized additive models (GAM) and spatially predictive maps of rockfish abundance at the
individual species level using habitat descriptors collected from visual surveys and fine-scale
bathymetry. We advanced beyond presence/absence and presence only models to create predic-
tive maps of fish density (100 m−2) and biomass (kg 100 m−2) for Sebastes rosaceus (rosy rockfish)
and S. constellatus (starry rockfish), both common species in commercial and recreational fish-
eries along the central coast of California. Selected models included co-variables of seafloor
depth, complexity, substratum type, and heterogeneity. Predicted density and biomass of both
species were highest in areas of complex rock on the continental shelf off Points Lobos and Sur at
50−90 (S. rosaceus) and 80−120 m (S. constellatus) water depth. Our results will be useful both in
stock assessments of these data-poor species as well as in allocation of fishing effort, catches, and
other space-based management decisions.
KEY WORDS: Visual surveys · Rockfishes · Generalized additive model · Central California
INTRODUCTION overfished by the National Marine Fisheries Service
and Pacific Fishery Management Council (PFMC
Pacific coast Sebastes spp. (rockfishes) are a wide- 2011). Rebuilding plans for overfished stocks are
ranging and diverse genus comprising at least 65 spe- required in accordance with the US Sustainable
cies that occupy most benthic habitats from shallow Fisheries Act of 1998 (16 U.S.C. § 1854[e][2]), which
estuaries and kelp forests to muddy slopes at 1500 m highlights the importance of identifying the physical
depth (Love et al. 2002). Several life-history character- and biological habitat factors that influence the dis-
istics (e.g. slow growth, longevity sometimes >100 yr, tribution and abundance of these fish species. A
and delayed sexual maturity; Love et al. 2002) make number of management actions have been taken to
most rockfish populations particularly vulnerable to support the rebuilding of rockfish stocks, including
high levels of fishing mortality. Approximately 40 spe- time and area closures, fishing gear modifications,
cies of rockfishes dominate deep-water fish assem- minimum stock size thresholds, and catch and size
blages in rocky habitats off California (Love & Yokla- limits. Timely stock assessments are needed to sup-
vich 2006) and support valuable recreational and port these management actions (Punt & Ralston
commercial fisheries that have been ongoing since 2007, Ralston & MacFarlane 2010), yet only 25 of
the mid-1800s (Love 2006, Miller et al. 2014). the 59 federally managed rockfish species have
Populations of 6 species of rockfishes currently been assessed (www.pcouncil.org/groundfish/stock-
are being rebuilt after having been classified as assessments/).
© The authors 2015. Open Access under Creative Commons by
*Corresponding author: lwedding@stanford.edu Attribution Licence. Use, distribution and reproduction are un-
restricted. Authors and original publication must be credited.
Publisher: Inter-Research · www.int-res.com
236 Mar Ecol Prog Ser 540: 235–250, 2015
Sedentary rockfishes living in heterogeneous, tively) (Love et al. 2002). Both species are commonly
high-relief, rocky habitats are difficult to appraise encountered in commercial and recreational fish-
accurately with conventional methods such as eries along the central coast of California (Love et al.
bottom-trawl gear. Fishery-independent surveys 2002) and yet are considered to be data-poor stocks
using non-extractive visual methods (e.g. from a (i.e. having only catch data available for assessments;
submersible) are an effective means to assess and Dick & MacCall 2010).
monitor stocks in rebuilding status and provide the Our overall goal was to model and map fish den-
opportunity to examine fish-habitat relationships at sity (100 m−2) and biomass (kg 100 m−2) of S.
finer spatial scales (1 to 5 m) than course (1 to 2 km) rosaceus and S. constellatus based on a variety of
benthic trawl gear (Anderson et al. 2009). For seafloor descriptors. The recent availability of
instance, habitat-specific estimates of density and detailed and accurate maps of substratum complex-
biomass from visual surveys (Yoklavich et al. 2007) ity from multibeam-acoustic surveys of the seafloor
are used in current stock assessments of S. levis within California’s territorial waters (i.e. those data
(cowcod). coming from the California Seafloor Mapping Pro-
Advances in seafloor mapping technologies, cou- gram) made it possible to produce regional maps of
pled with recent developments in modeling ap- predicted density and biomass at scales that are
proaches, support robust predictions of fish assem- ecologically relevant to these species. These results
blages (Moore et al. 2010, Pittman & Brown 2011) will improve our understanding of habitat variables
and individual species (Young et al. 2010). Rock- that influence the spatial distribution of these spe-
fishes demonstrate strong affinities to high- and low- cies, advance their stock assessments, and find
relief rocky substrata at specific depths (Yoklavich et application in the newly developed California Cur-
al. 2000, Laidig et al. 2009, Love et al. 2009). Addi- rent integrated ecosystem assessment (Levin &
tionally, habitat complexity and position relative to Schwing 2011).
the surrounding seabed have further informed
spatially predictive models of rockfish distributions
(Iampietro et al. 2008, Young et al. 2010). Predicting MATERIALS AND METHODS
rockfish density and biomass at the individual spe-
cies level, based on a variety of seafloor substratum Study area
variables, provides important information to improve
the assessment of fish stocks in rebuilding status. Our study area was located largely within state
In this study, we developed spatially predictive waters (3 nautical miles) off central California, USA,
models and maps of abundance for S. rosaceus (rosy including the southern portion of Monterey Bay,
rockfish) and S. constellatus (starry rockfish), typical Point Lobos, Point Sur, and Big Creek on the Big
members of the mid-shelf fish assemblage on rocky Sur coast (Fig. 1). We focused our research efforts in
substrata off central California, USA (Love & Yokla- depths from 35 to 150 m, bracketing the depths of
vich 2006). Rosy rockfish geographically range from occurrence for the 2 species of interest (Sebastes
the Strait of Juan de Fuca (Washington, USA) to rosaceus and S. constellatus), and across a range of
southern Baja California, Mexico, but are most abun- rocky habitats including extensive rock and boulder
dant from Cordell Bank off northern California to fields (e.g. off the Point Lobos and Point Sur head-
northern Baja California (Love 2011). Starry rockfish lands), isolated rocky outcrops and pinnacles that
range from northern California to southern Baja Cali- can be several meters high and surrounded by flat,
fornia, and are relatively abundant from central Cali- sandy seafloor, and on rock talus piles, scarps, and
fornia to southern Baja California. Juveniles and ledges in the heads of submarine canyons along the
adults of both species co-occur in the same depths Big Sur coast. Water is relatively cool and produc-
and substratum types. These are medium-size, tive along this section of the coast because the Cali-
mostly solitary species (maximum length 36 and fornia Current flows equatorward year round, and
46 cm for rosy and starry rockfish, respectively), and substantial upwelling of cold deep water occurs at
moderately long lived (likely maximum age at least the headlands, typically in spring and summer
401 and 32 yr for rosy and starry rockfish, respec- (Hickey 1998). Commercial and recreational fishing
with various types of gear has been supported by
1
Pers. comm., D. Pearson, Fisheries Ecology Division, South-
the diverse habitats on the continental shelf and
west Fisheries Science Center, NOAA, 110 Shaffer Rd., upper slope in this region for well over 60 yr (Love
Santa Cruz, CA 95060, USA. et al. 2002).
Wedding & Yoklavich; Rockfish predictive mapping 237
sors mounted on the outside of the submersible
recorded time, depth (pressure), and altitude at 1 s
intervals throughout each dive.
A total of 304 strip transects 2 m wide were con-
ducted for 10 min each along the seafloor. A distance
within 1 m of the seafloor and a speed of 0.5 to
1.0 knots were maintained during the transect sur-
veys. The submersible’s position was tracked at 1 to
3 s intervals using an ORE Trackpoint II ultra-short
baseline (USBL) acoustic system (EdgeTech) and
WINFROG software (v3.1; FUGRO). A scientific navi-
gator aboard the support vessel directed the start of
a transect once the submersible arrived at the pre-
determined target depth and location, and tracked
the submersible in real time within ArcGIS relative to
bathymetric maps. The length (m) of each transect
was estimated using a Doppler velocity log (DVL)
(NavQuest 600 Micro) and ring-laser gyrocompass
(CDL MiniRLG 2) attached to the outside of the sub-
mersible. Two video cameras (one outside and one
inside the submersible) were used to visually docu-
ment each transect.
A pilot operated the submersible while an experi-
enced scientist identified, counted, and estimated
total length (TL, cm; using paired lasers spaced
20 cm apart) of all fishes within a transect. Fish den-
sity (100 m−2) and biomass (kg 100 m−2) for Sebastes
rosaceus and S. constellatus were calculated from
Fig. 1. Delta submersible survey locations in Monterey Bay
each transect. To calculate biomass, TL of fish was
and along the Big Sur coast, off central California, USA
converted to weight (g) using W = 0.0052 × TL3.386
(R = 0.98259) for S. rosaceus and W = 0.0097 × TL3.160
Data sets (R = 0.97805) for S. constellatus (Love et al. 1990).
The relationship between length and weight was
Visual surveys of demersal fishes and components identical for males and females for both species.
of their habitats were conducted using a manned Seafloor substratum types (mud, sand, gravel,
submersible (Delta) during daytime hours from Sep- pebble, cobble, boulder, continuous flat rock, rock
tember to November in 2007 and 2008 (see Yoklavich ridge and pinnacle top) were classified from the
et al. 2000, 2002, 2007 for details on the Delta survey video of each transect, based on geological defini-
vehicle and methods). Dives were positioned ran- tions detailed in Greene et al. (1999). Distinct habitat
domly in areas of rocky substrata that were identified patches, with a minimum duration of 3 s, were de-
from maps of bathymetry (Monterey Bay Aquarium fined by assigning primary (at least 50% of the area
Research Institute Mapping Team, Monterey Bay viewed) and secondary (> 20% of the area viewed)
Multibeam Survey; Seafloor Mapping Lab, Califor- substratum types each time a change was noted
nia State University Monterey Bay [SFML-CSUMB]) along the transect.
and from interpreted seafloor habitats (Yoklavich et The length of each patch was determined from
al. 1997, Eittreim et al. 2000). Surveys were con- the DVL measurements. Primary substrata were
ducted by traversing haphazardly across substratum grouped based on levels of structural complexity
types and depth gradients within designated rocky (Table 1). Pinnacle top, rock, and flat rock were clas-
habitats. Soft sediment was not specifically targeted, sified as high complexity or large structured hard
but was surveyed along with rocky substrata on the substratum type (Lhard). Boulders, cobble, pebble,
quantitative sample transects. The same general and gravel were classified as medium complexity
areas were surveyed in 2007 and 2008, but transects hard substratum (Mhard). Sand and mud comprised
were not re-sampled from one year to the next. Sen- a soft substratum group (Soft) of low complexity.
238 Mar Ecol Prog Ser 540: 235–250, 2015
Table 1. Habitat characterization groups from submersible Following our visual surveys, high-resolution multi-
survey video analysis beam data were collected throughout the entire
study area by the California Seafloor Mapping Pro-
Habitat image Habitat type and complexity gram (http://walrus.wr.usgs.gov/mapping/csmp). A
5 m resolution bathymetric Digital Elevation Model
Habitat code: Lhard
(DEM) and a derived Rough/Smooth classification of
Habitat description:
pinnacle top, rock, flat rock the seafloor were provided in grid format. Slope-of-
Level of complexity: high
slope (i.e. habitat complexity) was derived from the
bathymetric grids using the ArcGIS Spatial Analyst
extension (ESRI) slope function to obtain the rate of
Habitat code: Mhard change of slope. Bathymetric position index (BPI)
Habitat description: was calculated using the Benthic Terrain Modeler
boulder, cobble, pebble, extension in ArcGIS. BPI is a second order derivative
gravel of the multibeam bathymetry data and characterizes
Level of complexity: medium a pixel in the bathymetric DEM as a positive (e.g.
pinnacle top) or negative (e.g. canyon) feature of the
Habitat code: Soft surrounding seascape (Lundblad et al. 2006, Young
Habitat description: et al. 2010). BPI grid creation and classification was
sand and mud applied using a scales of analysis at 5 pixel (25 m)
Level of complexity: low and 50 pixel (250 m) annulus thickness for fine and
broad respectively. Finally, the covariates were ex-
tracted for each transect using a 100 m radius moving
window analysis in GIS.
An entire transect was characterized as a weighted
sum of the patch substrata (O’Farrell et al. 2009). For Statistical analysis
substratum type b (b ∈ Lhard, Mhard, Soft), in tran-
sect t, the numerical value describing that substra- Predictive models of fish density (100 m−2) and bio-
tum’s contribution (S) was computed as: mass (kg 100 m−2) of Sebastes rosaceus and S. con-
stellatus were developed using generalized additive
⎡ Lt , p ⎤
Sb ,t = ∑ ⎢Sb ,t , p × (1)
∑ P Lt , p ⎥⎦
models (GAM; Pinheiro & Bates 2009, Zuur et al.
P ⎣
2009). These species are known to occur at specific
where Lt,p represents the length of patch p in transect depths and on distinct substratum types (Love et al.
t. The resulting transect length-weighted habitat 2002); we used the flexible GAM to accommodate
value characterized the 3 main substratum types (e.g. the expected non-linear responses of both species to
Lhard, Mhard, Soft) for each transect. This value, our co-variates. Two sets of models were developed:
ranging from 0 (low complexity) to 0.7 (high com- (1) models based on the co-variables depth, a tran-
plexity), was used as a metric of habitat complexity sect length-weighted value for the 3 substratum
for each visual transect. Habitat heterogeneity was groups (i.e. Lhard, Mhard, Soft), and habitat hetero-
estimated as number of habitat patches per transect geneity, which were collected during the visual sur-
(Table 2). veys, and measures of seafloor roughness, habitat
Table 2. Environmental co-variables collected during visual surveys and used in generalized additive models (GAMs) of rockfish density
and biomass
Co-variable Description Units Source Range Mean (SD)
Depth Depth at start of transect meters Submersible sensor 25−315 137.4 (83.1)
Lhard* Pinnacle top, rock, flat rock length-weighted habitat value Sb,t = ∑P [Sb,t,p × Lt,p / ∑PLt,p] 0−0.7 0.4 (0.2)
Mhard* Boulder, cobble, pebble, gravel length-weighted habitat value Sb,t = ∑P [Sb,t,p × Lt,p / ∑PLt,p] 0−0.7 0.1 (0.1)
Habitat Measure of seascape patchiness # habitat patches on transect line Video analysis 1−57 18.2 (9.8)
heterogeneity
*The length-weighted habitat value was calculated using Sb,t = ∑P [Sb,t,p × Lt,p / ∑PLt,p] based on O’Farrell et al. (2009)
Wedding & Yoklavich; Rockfish predictive mapping 239
complexity, and BPI derived from the high-resolution rock, and flat rock), 15% Mhard (boulder, cobble,
multibeam bathymetry; and (2) models based only pebble, and gravel), and 15% Soft (sand and mud)
on the co-variables derived from the multibeam substrata at depths of 35 to 150 m. Average density
bathymetry. Calculations were computed using R for S. rosaceus and S. constellatus was 3.42 (SE =
software and the mgcv package (Wood 2004) with a 0.16) and 0.74 (0.06) fish 100 m−2, respectively. On
Gaussian error distribution and an identity link. Data transects, the average biomass was 0.26 kg 100 m−2
exploration followed protocols described by Zuur et (SE = 0.01) for S. rosaceus, and 0.14 kg 100 m−2 (0.01)
al. (2010). Cleveland dotplots and boxplots were for S. constellatus. Density of both species was rela-
used to determine the presence of outliers. Collinear- tively high on transects at Carmel Canyon, Point Sur,
ity was investigated between covariates using Pear- and Sur Canyon sites (Figs. 2a & 3a). Spatial pattern
son correlation coefficients, multi-panel scatterplots, of S. rosaceus biomass was similar to that of density
and variance inflation factors (VIF) (Montgomery & along the Central California coast, with higher bio-
Peck 1992). mass coincident with areas of greater habitat struc-
Akaike’s information criterion (ΔAIC) and Akaike tural complexity off the Monterey Peninsula, Point
weights (wi ) were used to select the most parsimo- Lobos, and Point Sur (Fig. 2d). There were several
nious models among all possible covariate combina- pockets of relatively high biomass of S. constellatus
tions (Burnham & Anderson 2002). We selected mod- along the central Californian coast, from the Mon-
els based on lowest AIC values using the MuMIn terey Peninsula to Big Creek, the most southern
package (Barto 2011) and models having ΔAIC < 2 reaches of our study area (Fig. 3d).
were combined using a weighted (wi ) model aver-
age. We used k fold cross validation in which we split
the data into equal sized parts and then iteratively Modeling of Sebastes rosaceus density and biomass
used part of the data to fit the model and a different
part to test it (Hastie et al. 2009). We repeated each The GAMs to predict fish density and biomass
k-fold cross validation process 500 times and exam- were based on habitat co-variables from in situ visual
ined the distribution of coefficient of determination. surveys and multibeam acoustic bathymetry.
Validation of the optimum model was accomplished The selected GAM to predict S. rosaceus density
by inspecting homogeneity (plotting residual vs fit- included depth and hard rock substrata with high
ted values) and independence (variogram of residu- and medium complexity (wi = 0.99; Table 3). This
als and plotting residuals versus each covariate). selected model accounted for 42% of the variability
Spatially predictive mapping was conducted in in S. rosaceus density and had a mean R2 = 0.40 for
ArcGIS 10.1 using the Marine Geospatial Ecology 500 runs of 10-fold cross-validation. From response
Tool (MGET; v0.8a48) (Roberts et al. 2010), which curves, S. rosaceus density increased with depth
integrates ArcGIS with the R statistical package (R from 35 to ~70 m and then declined relatively steeply
Development Core Team 2011). In MGET, the GAM in deeper waters (Fig. 4a). The response curves for
tool was applied to derive spatially predictive rasters transect length-weighted habitat values indicated a
for rockfish density and biomass. Prior to predictive steady increase in predicted S. rosaceus density with
mapping of the data for this analysis, data were ran- increasing complexity for both large (pinnacle top
domly split into ‘test’ (1/3) and ‘training’ (2/3) sets to and rock habitat, Fig. 4b) and medium-to-small
assess map accuracy. The training data sets were substrata (boulder, cobble, pebble, gravel habitat,
used to create the predictive maps, and the test data Fig. 4c).
sets were applied to evaluate map accuracy assess- The selected GAM to predict S. rosaceus biomass
ment using Kendall’s τ and Spearman’s ρ. included depth, hard rock substrata with high and
medium complexity, habitat heterogeneity (e.g.
number of patches per transect), and BPI (wi = 0.81;
RESULTS Table 3), and accounted for 46% of the biomass vari-
ability (mean R2 = 0.42). The response curves for
Spatial patterns of rockfish density and biomass depth and both transect length-weighted habitat val-
ues (Lhard and Mhard) revealed the same patterns as
Sebastes rosaceus and S. constellatus density and in the density response curves (Fig. 4d−f). From the
biomass data were used in our modeling efforts BPI response curve (Fig. 4g), biomass was lowest in
based on a total of 304 transects covering 146 000 m2 canyons and depressions (negative values of BPI)
of seafloor that comprised 70% LHard (pinnacle top, and demonstrated an overall increase to high BPI
240 Mar Ecol Prog Ser 540: 235–250, 2015
Fig. 2. Sebastes rosaceus. Density and biomass off central California, USA: (a) graduated dots of observed density; (b) pre-
dicted density throughout study area; (c) predicted density in enlarged map of area off Pt. Lobos; (d) graduated dots of ob-
served biomass; (e) predicted biomass throughout study area; and (f) predicted biomass in enlarged map of area off Pt. Lobos
Wedding & Yoklavich; Rockfish predictive mapping 241
Fig. 3. Sebastes constellatus. Density and biomass off central California, USA: (a) graduated dots of observed density; (b) pre-
dicted density throughout the study area; (c) predicted density in enlarged map of area off Pt. Sur; (d) graduated dots of ob-
served biomass; (e) predicted biomass throughout the study area; and (f) predicted biomass in enlarged map off Pt. Sur
242 Mar Ecol Prog Ser 540: 235–250, 2015
Table 3. Results of model selection to predict Sebastes rosaceus density and biomass off central California for 2 sets of models:
(1) visual and bathymetry models and (2) bathymetry alone to support GIS-based predictive mapping. Models were ranked by
Akaike’s information criterion (ΔAIC) and Akaike weights (wi ). Models having ΔAIC < 2 were combined using a weighted (wi )
model average and are in bold. Lhard = high complexity or large structured hard substratum (pinnacle top, rock, and flat
rock); Mhard = medium complexity hard substratum (boulders, cobble, pebble, and gravel); Soft = low complexity substratum
(sand and mud)
Model type Model selection results ΔAIC wi
Models based on co-variables from both visual surveys and multibeam bathymetry
Density S. rosaceus ~ Depth + Lhard + Mhard 0 0.99
Density S. rosaceus ~ BPI + Lhard + Mhard 28.65 0.01
Biomass S. rosaceus ~ Depth + Lhard + Mhard + Habitat Heterogeneity + BPI 0 0.81
Biomass S. rosaceus ~ Depth + Lhard + Mhard + BPI 3.71 0.13
Biomass S. rosaceus ~ Depth + Lhard + Mhard + BPI + factor(Rough) 5.39 0.01
Models based on co-variables derived from multibeam bathymetry
Density S. rosaceus ~ Depth + BPI 0 0.45
Density S. rosaceus ~ Depth + BPI + factor(Rough) 0.96 0.28
Density S. rosaceus ~ Depth 2.12 0.15
Density S. rosaceus ~ Depth + factor(Rough) 2.49 0.13
Biomass S. rosaceus ~ Depth + BPI + Habitat Complexity 0 0.62
Biomass S. rosaceus ~ Depth + BPI + Habitat Complexity + factor(Rough) 1.69 0.27
Biomass S. rosaceus ~ Depth + BPI + factor(Rough) 4.41 0.07
features (e.g. pinnacle tops). Biomass of S. rosaceus Complexity + factor (Rough), wi = 0.27 (Table 3). This
was highest in areas of ~4 to 6 habitat patches per model accounted for 38% of the biomass variability
100 m2 (Fig. 4h), with declining biomass in areas (R2 = 0.35). Response curves for depth (Fig. 5c), BPI
of very low and high habitat heterogeneity (e.g. (Fig. 5d), and habitat complexity (Fig. 5e) had greater
patchiness). biomass values for the rough substratum factor than
for the smooth substratum factor. The response curve
for predicted biomass versus depth was similar to
Predictive maps of Sebastes rosaceus density that of the density curve, and demonstrated an
and biomass increase in S. rosaceus biomass from 35 to ~70 m and
a decrease in biomass at greater depth ranges
The predictive maps for density and biomass were (Fig. 5c). In the response curve for BPI, S. rosaceus
based solely on co-variates derived from acoustic biomass followed the same pattern as the density
multibeam bathymetry. curve and was lowest in canyons and depressions
The selected GAM to create a predictive map of S. and was characterized by distinct peaks in biomass
rosaceus density accounted for 33% of its density associated with low-relief habitat features (e.g. boul-
variability (R2 = 0.32). The model was based on a der fields) and high BPI features (e.g. pinnacle tops)
weighted average of M1: S. rosaceus ~ Depth + BPI, (Fig. 5d). Habitat complexity was not important for
wi = 0.45 and M2: S. rosaceus ~ Depth + BPI + fac- modeling density, but for biomass this response
tor(Rough), wi = 0.28 (Table 3). Model-averaged re- curve indicated a sustained increase in S. rosaceus
sponse curves for depth (Fig. 5a) and BPI (Fig. 5b) biomass with increasing complexity (Fig. 5e).
had greater density values for the rough substratum Predictive maps of S. rosaceus density (Fig. 2b) and
factor than for the smooth factor. In the response biomass (Fig. 2e) were produced based on the aver-
curve for BPI, density was lowest in canyons and age of the top 2 models by weight. Spatial patterns of
depressions (similar to Fig. 4g), but also demon- biomass predicted across the study region were sim-
strated distinct peaks in density associated with ilar to that of density, with higher biomass coincident
medium-relief habitat features (e.g. boulder fields) with areas of greater habitat structural complexity off
and high BPI features (e.g. pinnacle tops) (Fig. 5b). the Monterey Peninsula, Point Lobos (Fig. 2c,f), and
The selected GAM for S. rosaceus biomass predic- Point Sur. The highest predicted density and biomass
tive map was a weighted model average M1: S. of S. rosaceus rockfish were mapped on the continen-
rosaceus ~ Depth + BPI + Habitat Complexity, wi = tal shelf at 50 to 90 m water depth off Point Pinos and
0.62 and M2: S. rosaceus ~ Depth + BPI + Habitat Point Sur. The predictive maps of S. rosaceus density
Wedding & Yoklavich; Rockfish predictive mapping 243
Fig. 4. Sebastes rosaceus. Response curves, based on habitat co-variables from visual surveys and acoustic multibeam bathy-
metry, for generalized additive model (GAM) predicted density versus (a) depth, (b) length-weighted habitat values in high-
relief rock (Lhard) and (c) length-weighted habitat values in low-relief rock (Mhard); and GAM-predicted biomass versus
(d) depth, (e) length-weighted habitat values in Lhard, (f) length-weighted habitat values in Mhard, (g) bathymetric position
index (BPI), and (h) habitat heterogeneity. Solid lines = mean (±1 SE, dashed lines). Rug plots along the x-axis = calibration
data points
and biomass had an accuracy of Spearman’s ρ = 0.63 constellatus density (R2 = 0.41). From response
and 0.66 and Kendall’s τ = 0.46 and 0.49, respectively. curves, S. constellatus occurred deeper than S.
rosaceus, with relatively high densities in water
depths > 80 m (Fig. 6a). Similar to S. rosaceus, the
Modeling of Sebastes constellatus density response curves for transect length-weighted habitat
and biomass values indicated a steady increase in predicted S.
constellatus density with increasing complexity for
The selected GAM to predict S. constellatus den- both large (pinnacle top and rock habitat, Fig. 6b)
sity included depth and hard rock substrata with and medium-to-small rocky substrata (boulder, cob-
high and medium complexity (wi = 0.99, Table 4), and ble, pebble, gravel habitat, Fig. 6c).
concluded in similar results to the S. rosaceus model. The model for S. constellatus biomass was a
The model accounted for 43% of the variability in S. weighted average of depth, hard rock substrata
244 Mar Ecol Prog Ser 540: 235–250, 2015
Fig. 5. Sebastes rosaceus. Response curves, based on habitat
co-variables derived from acoustic multibeam bathymetric
surveys, for generalized additive model (GAM) predicted
density versus (a) depth and (b) bathymetric position index
(BPI); and GAM-predicted biomass versus (c) depth, (d) BPI,
and (e) habitat complexity with factor representing rough
substratum (red lines) and smooth substratum (blue lines).
Solid lines = mean (±1 SE, dashed lines). Rug plots along the
x-axis = calibration data points
with high and medium complexity, habitat hetero- BPI) and peaked at low-relief habitat features (e.g.
geneity, and BPI (Table 4), which accounted for boulder fields, Fig. 6g). The response curve for
54% of its variability (R2 = 0.40). The response habitat heterogeneity (Fig. 6h) demonstrated a dif-
curves for depth (Fig. 6d) and one transect length- ferent pattern than that of S. rosaceus biomass; pre-
weighted habitat co-variable (i.e. Mhard; Fig. 6f) dicted biomass of S. constellatus was greater in
revealed the same patterns as in density response areas of more homogenous habitat (i.e. low numbers
curves for this species. However, the response curve of habitat patches per m2).
for the transect length-weighted habitat co-variable
Lhard (Fig. 6e), which characterizes highly complex
pinnacle top and rock, demonstrated an increase in Predictive maps of Sebastes constellatus density
biomass to an intermediate level followed by a and biomass
steady increase at the highest complexity. As with
S. rosaceus, BPI was not an important co-variate in The selected GAM to predict S. constellatus den-
the model that predicted density but was included sity was a weighted average of depth, BPI and rough
in the selected GAM to predict biomass for S. con- substratum as a factor (Table 4). This model ac-
stellatus; predicted biomass was lowest in subma- counted for 29% of the variability in predicted S. con-
rine canyons and depressions (negative values of stellatus density (R2 = 0.27). Density of S. constellatus
Wedding & Yoklavich; Rockfish predictive mapping 245
Table 4. Results of model selection to predict Sebastes constellatus density and biomass off central California for 2 sets of
models: (1) visual and bathymetry models and (2) bathymetry alone to support GIS-based predictive mapping. Models are
ranked by Akaike’s information criterion (ΔAIC) and Akaike weights (wi ). Models having ΔAIC < 2 were combined using a
weighted (wi ) model average and are in bold. Lhard = high complexity or large structured hard substratum (pinnacle top,
rock, and flat); Mhard = medium complexity hard substratum (boulders, cobble, pebble, and gravel); Soft = low complexity
substratum (sand and mud)
Model type Model selection results ΔAIC wi
Models based on co-variables from both visual surveys and multibeam bathymetry
Density S. constellatus ~ Depth +Lhard +Mhard 0 0.99
Density S. constellatus ~ BPI +Lhard + Depth 33.01 0.01
Biomass S. constellatus ~ Depth + Lhard +Mhard+ Habitat Heterogeneity + BPI 0 0.60
Biomass S. constellatus ~ Depth +Lhard +Mhard + BPI 1.48 0.29
Biomass S. constellatus ~ Depth +Lhard +Mhard + BPI + factor(Rough) 3.45 0.10
Models based on co-variables derived from multibeam bathymetry
Density S. constellatus ~ Depth + BPI 0 0.72
Density S. constellatus ~ Depth + BPI + factor(Rough) 1.94 0.27
Density S. constellatus ~ Depth 23.14 0.01
Biomass S. constellatus ~ Depth + BPI + Habitat Complexity 0 0.54
Biomass S. constellatus ~ Depth + BPI + Habitat Complexity + factor(Rough) 0.36 0.45
Biomass S. constellatus ~ Depth + Habitat Complexity 7.71 0.01
peaked at 70 to 90 m in the response curves for depth Spearman’s ρ = 0.36 and 0.63 and Kendall’s τ = 0.53
of both substratum types (Fig. 7a). The response and 0.46, respectively. Spatial patterns of predicted
curve for BPI indicated that S. constellatus density biomass of S. constellatus were highest in areas of
was lowest in canyons and depressions followed by a complex rock off Point Lobos (Fig. 3e) and Point Sur
steep increase to relatively low positive BPI values (Fig. 3e,f); these areas were also associated with high
that are associated with fine-scale habitat features predicted biomass of S. rosaceus.
such as boulder or rocky substrata at higher elevation
relative to surrounding seascape (Fig. 7b). Density
increased even more at the highest BPI values, indi- DISCUSSION
cating an association of S. constellatus with pinnacle
tops. Until recently, predictive habitat-based models and
The selected GAM for S. constellatus biomass was maps of demersal marine fish distribution had been
a weighted average of depth, BPI, habitat complex- largely developed from presence/absence or pres-
ity, and rough substratum as a factor (Table 4). This ence-only fish data (Iampietro et al. 2008, Young
model accounted for 32% of the variability in S. con- et al. 2010). In this paper, we attempt to advance
stellatus density (R2 = 0.23). From response curves of beyond presence/absence and presence only models
co-variates depth (Fig. 7c), BPI (Fig. 7d), and habitat to develop predictive regional maps of rockfish den-
complexity (Fig. 7e), S. constellatus biomass was sity and biomass at the individual species level in a
greater for the rough substratum factor than for the temperate ecosystem. There have been a number of
smooth factor. Peak biomass of S. constellatus was spatially predictive mapping and modeling studies in
predicted at deeper depth (i.e. ~100 m) than that of tropical ecosystems. Costa et al. (2014) integrated
peak density (~80 m). The response curve for BPI habitat data from acoustic sensors (i.e. splitbeam and
indicated that S. constellatus biomass was lowest in multibeam echosounders) in the US Virgin Islands to
canyons and depressions followed by a sharp peak predict fish density, and found the model performed
associated with fine-scale habitat features that rise best at larger body sizes (≥ 29 cm) to identify fish
slightly above the surrounding seascape (Fig. 7d). aggregations and help coastal managers prioritize
The habitat complexity response curve indicated an areas of higher conservation value. Pittman & Brown
increase in S. constellatus biomass with increasing (2011) similarly developed predictive maps for sev-
habitat complexity. eral key fish species associated with Caribbean coral
The accuracies of the predictive maps of S. constel- reef seascapes and found that habitat complexity
latus density and biomass were characterized by derived from Light Detection and Ranging Data
246 Mar Ecol Prog Ser 540: 235–250, 2015
Fig. 6. Sebastes constellatus. Response curves, based on habitat co-variables from visual surveys and acoustic multibeam
bathymetry, for generalized additive model (GAM) predicted density versus (a) depth, (b) length-weighted habitat values in
high-relief rock (Lhard) and (c) length-weighted habitat values in low-relief rock (Mhard); and GAM-predicted biomass
versus (d) depth, (e) length-weighted habitat values in Lhard, (f) length-weighted habitat values in Mhard, (g) bathymetric
position index (BPI), and (h) habitat heterogeneity. Solid lines = mean (±1 SE, dashed lines). Rug plots along the x-axis =
calibration data points
(LiDAR) contributed most to the spatial model of probability of occurrence (presence/absence) of Se-
habitat suitability for Stegastes planifrons (threespot bastes rosaceus in high relief rocky areas of Cordell
damselfish). Habitat complexity, particularly the Bank on California’s northern coast, with habitat
slope of the slope (a measure of the maximum rate of complexity included as a strong predictor in these
slope change) was found to be the most useful pre- models. We found densities of S. rosaceus and S. con-
dictor of diversity and abundance of fishes and corals stellatus to have similar patterns of habitat affinity in
in the Caribbean (Pittman et al. 2009). relation to fine-scale remotely sensed measures of
Many rockfish species demonstrate habitat prefer- habitat complexity. We also found that intermediate
ences for complex rocky substrata (Love & Yoklavich levels of habitat heterogeneity were important in
2006, Love et al. 2009). Young et al. (2010) applied explaining S. rosaceus variability across the seascape
generalized linear models to predict the highest and to demonstrate the importance of composition
Wedding & Yoklavich; Rockfish predictive mapping 247
Fig. 7. Sebastes constellatus. Response curves, based on
habitat co-variables derived from acoustic multibeam bathy-
metric surveys, for generalized additive model (GAM) pre-
dicted density versus (a) depth and (b) bathymetric position
index (BPI); and GAM-predicted biomass versus (c) depth,
(d) BPI, and (e) habitat complexity with factor representing
rough substratum (red lines) and smooth substratum (blue
lines). Solid lines = mean (±1 SE, dashed lines). Rug plots
along the x-axis = calibration data points
and configuration of habitat features. Further, the cally evaluated several of these assumptions using
BPI co-variable highlights areas in the seascape that similar survey techniques in prior field surveys
are tops of large pinnacles or rocky outcrop features (Yoklavich et al. 2007, Laidig et al. 2013). The strip
adjacent to canyons (e.g. Point Pinos and Point Sur). transect method used in our study assumes 100%
Anderson & Yoklavich (2007) reported S. rosaceus detection of the target species within the strip. To
rockfishes were found in groups of other large- help meet this assumption, we used a relatively nar-
bodied rockfishes (e.g. S. paucispinis, S. flavidus, S. row strip width (2 m) during our surveys. That said, it
rubrivinctus, and Sebastomus spp.) with strong asso- is unlikely that 100% of S. rosaceus and S. constel-
ciations to high-relief rocky outcrops. In addition to latus were seen in the transects, particularly the
the refuge provided by structurally complex habitat, smallest individuals nestled in the rocky substrata.
relatively productive waters associated with rocky This would result in an underestimation of densities
pinnacles and outcrops in and adjacent to canyons by some unknown amount. Additional studies will be
could support greater fish biomass. required to estimate true detectability of these spe-
The visual survey methods used in this study can cies in high-relief habitats.
be more effective than extractive trawl surveys in Another important assumption of these underwater
estimating abundance of rockfish species living in surveys is that rockfish behavior is independent of
high-relief rocky areas. As with all survey methods, the observer and submersible (i.e. no avoidance or
there are several assumptions and sources of uncer- attraction). Laidig et al. (2013) reported that 6 and
tainty associated with visual surveys. We have criti- 10% of S. rosaceus near (total n = 134) and on (n = 10)
248 Mar Ecol Prog Ser 540: 235–250, 2015
the seafloor, respectively, reacted mostly by swim- geographic scale of our study could also be expanded
ming toward or to the left of the Delta submersible to include data from our visual surveys in rocky areas
during surveys off central California. Sebastes con- of southern California where both target species are
stellatus reacted similarly, with 5 and 8% of fish near common.
(n = 21) and on (n = 13) the seafloor responding to the We developed 2 sets of models in this study: one set
survey vehicle2. Both species moved an average dis- based on co-variables from both the visual surveys
tance of 2 to 3 m per reaction. This type of reaction and region-wide high-resolution bathymetry and the
could bias estimated densities if the fish entered or other set of models based only on co-variables de-
exited the strip transect; otherwise, these relatively rived from the bathymetry. Our motivation was to
small movements would not influence the resultant examine the added value of data collected in situ
densities. In addition, the assumption that the fish are from a submersible compared to that collected solely
distributed randomly with respect to the transect was from acoustic surveys. The models using all of the co-
met by randomizing survey sites and traversing hap- variates (including those from the visual surveys)
hazardly across substratum types and depth gradi- accounted for more of the overall variance (42 to
ents within designated rocky habitats. 54%) in estimated density and biomass for both spe-
There also are potential sources of error related to cies than those using only derived variables from
fish measurements. In an earlier study (Yoklavich et multibeam bathymetry (29 to 38%). Clearly data
al. 2007), we estimated error associated with our from the visual surveys improved the predictive
underwater estimates of fish size. From the sub- capabilities of the models. However, in order to apply
mersible we measured fish replicas of known total predictions on a region-wide scale, only the bathy-
length, and size generally was underestimated by a metric co-variates could be used because in situ data
relatively small amount (mean ± SD deviation: −1.1 ± were not available on a broad scale.
1.2 cm). This would result in an underestimate of bio- Benthic terrain analysis of multibeam bathymetric
mass. Biomass estimates also contain some unknown acoustic data is a valuable way to identify seafloor
amount of error related to the conversion of length to habitat that supports individual species and assem-
weight. However, the regression of weight and blages (Wilson et al. 2007, Guinan et al. 2009). Quan-
length (Love et al. 1990) provided an excellent fit for tifying and mapping elements of rockfish habitat,
both of our target species, so the amount of error such as seafloor substratum type, texture, and com-
introduced using length to calculate weight is ex- plexity, are critical for evaluating the effectiveness of
pected to be small. these areas to maintain rockfish stocks (Yoklavich et
Further, our research is limited by the temporal and al. 2000, 2007). Our ability to derive indices of ben-
spatial extent of the study. Our visual surveys were thic habitats (e.g. habitat complexity, depth, BPI)
conducted only in Fall (September to November) of from multibeam acoustic data3,4 that were recently
2007−2008, and we did not expect a seasonal effect synthesized across California state waters supports
in abundance because these sedentary rockfish spe- the production of spatially predictive maps of demer-
cies are not known to be wide-ranging (Love et al. sal fish populations at a regional scale relevant to the
2002). However, repeating these surveys to establish assessment of fish stocks in rebuilding status. Fur-
time series in density and biomass over several years ther, providing a spatial component to these predic-
would allow us to evaluate change in rockfish abun- tions can be critical in the management of relatively
dance across the region as well as inside and outside sedentary rockfish species to safeguard against local
marine protected areas (MPA). In particular, this depletion (Parker et al. 2000).
study represents a baseline for monitoring the deep- Currently it is challenging to integrate oceano-
water portion of 8 MPAs that were established on the graphic and benthic habitat predictor variables into
central coast coincident with the commencement of habitat models and maps, largely because the spatial
our surveys in September 2007. With future monitor- resolution of available oceanographic data (e.g. tem-
ing, we will be able to include an MPA term in the perature, salinity, and bottom currents from regional
models to evaluate the efficacy of these closed areas oceanographic modeling systems [ROMS]) is 10s of
in protecting deep-water species of rockfishes. The km, and the resolution of the benthic habitat data is
several orders of magnitude greater (<1 m from
2
Pers. comm., T. Laidig, Fisheries Ecology Division, South-
3
west Fisheries Science Center, NOAA, 110 Shaffer Rd., http://walrus.wr.usgs.gov/mapping/csmp/
4
Santa Cruz, CA 95060, USA http://seafloor.otterlabs.org/csmp/csmp.html
Wedding & Yoklavich; Rockfish predictive mapping 249
visual surveys and 2−5 m from multibeam acoustic Acknowledgements. We thank R. Starr, T. Laidig, M. Love,
surveys of bathymetry). Our models were developed M. Nishimoto, L. Snook and others for help with field data
collection. We are especially grateful to T. Laidig for post-
only with benthic habitat predictors. However, the cruise video analysis and to D. Watters for database man-
inclusion of oceanographic variables could further agement. We thank D. Huff for his expertise and help with
improve the predictive capability of our models and data analyses. J. Blakley assisted with derived seafloor
maps when these data become available at finer indices. The in situ visual survey data were collected as part
of a grant from the California Ocean Protection Council to R.
spatial scales.
Starr and M.M.Y., and the analytical work was supported by
Rockfishes are among the most valuable fisheries a grant from NOAA NMFS Habitat Assessment Improve-
in California, have extremely vulnerable life-history ment Plan Program to M.M.Y. Acoustic data used in this
characteristics, and cannot sustain levels of fishing study were acquired, processed, archived, and distributed
by the Seafloor Mapping Lab of California State University
mortality; as a result, they are being managed more
Monterey Bay, supported in part by California State Map-
conservatively than in the past. The predictive maps ping Program. We appreciate the very thoughtful and useful
of density and biomass of S. rosaceus and S. constel- comments and suggestions from E. J. Dick, D. Huff, S. Sog-
latus will improve our understanding of habitat vari- ard, and 3 anonymous reviewers.
ables that influence the spatial distribution and
abundance of these species across the central coast. LITERATURE CITED
The results of our study provide information to sup-
port stock assessments of these 2 species of rock- Anderson TJ, Yoklavich MM (2007) Multiscale habitat asso-
fishes, both considered to be data-poor species by ciations of deepwater demersal fishes off central Califor-
nia. Fish Bull 105:168−179
the PFMC. For example, using our predictive maps
of biomass, stock assessors could estimate total bio-
➤ Anderson TJ, Syms C, Roberts DA, Howard DF (2009) Multi-
scale fish-habitat associations and the use of habitat surro-
mass (and associated uncertainty) of these species gates to predict the organization and abundance of deep-
within our designated study area off central Califor- water fish assemblages. J Exp Mar Biol Ecol 379:34−42
nia and consider those estimates in context with Barto K (2011) MuMl package: multi-model inference.
www.cran.r-project.org/web/packages/
assessments based solely on commercial and recre- Burnham KP, Anderson DR (2002) Model selection and
ational landings (Dick & MacCall 2010, Ralston et multimodel inference, 2nd edn. Springer-Verlag, New
al. 2010). Such estimates of habitat-specific abun- York, NY
dance are critical to effectively control the magni- ➤ Costa B, Taylor JC, Kracker L, Battista T, Pittman S (2014)
Mapping reef fish and the seascape: using acoustics and
tude of fishing mortality and to support the local spatial modeling to guide coastal management. PLoS
recovery of depleted populations. ONE 9:e85555
Dick EJ, MacCall AD (2010) Estimates of sustainable yield
for 50 data-poor stocks in the Pacific coast groundfish
fishery management plan. NOAA Tech Memo NOAA-
CONCLUSIONS
TM-NM FS-SWFSC-460
Eittreim SL, Anima RJ, Stevenson AJ, Wong FL (2000)
In this study we developed models of rockfish den- Seafloor rocks and sediments of the continental shelf
sity and biomass based on in situ observations of fish from Monterey Bay to Point Sur, California. USGS
numbers, size, and associated seafloor substratum Miscellaneous Field Studies Map MF-2345, online v1.0.
http://pubs.usgs.gov/mf/2000/2345/
variables and derived variables from region-wide
➤ Greene HG, Yoklavich MM, Starr RM, O’Connell VM and
multibeam bathymetry. Model results were expres- others (1999) A classification scheme for deep seafloor
sed as predictive maps of density and biomass on a habitats. Oceanol Acta 22:663−678
regional scale. Such maps allow us to quantify habi- ➤ Guinan J, Brown C, Dolan MF, Grehan AJ (2009) Ecological
niche modelling of the distribution of cold-water coral
tat capacity, prioritize habitat conservation, and eva-
habitat using underwater remote sensing data. Ecol
luate potential risk of various human activities to Inform 4:83−92
rockfish populations over broad spatial scales. These Hastie T, Tibshirani R, Friedman JH (2009) The elements of
results have direct application to coastal and marine statistical learning: data mining, inference, and predic-
tion. Springer series in statistics. Springer, New York, NY
spatial management, particularly in (1) the design
Hickey BM (1998) Coastal oceanography of western North
and monitoring of MPAs, (2) improving the identifi- America from the tip of Baja California to Vancouver
cation of essential fish habitats, (3) distinguishing Island. In: Robinson AR & Brink KH (eds) The Sea, Vol
those areas important to the restoration or rebuilding 11. The global coastal ocean: processes and methods.
of depleted stocks, and (4) advancing our under- John Wiley & Sons, New York, NY, p 345−393
standing of the effects of a changing climate on the ➤
Iampietro PJ, Young MA, Kvitek RG (2008) Multivariate
prediction of rockfish habitat suitability in Cordell Bank
spatial distribution and abundance of Pacific coast National Marine Sanctuary and Del Monte Shalebeds,
groundfishes. California, USA. Mar Geod 31:359−371
250 Mar Ecol Prog Ser 540: 235–250, 2015
➤ Laidig TE, Watters DL, Yoklavich MM (2009) Demersal fish S-PLUS. Springer Verlag, New York, NY
and habitat associations from visual surveys on the cen- ➤ Pittman SJ, Brown KA (2011) Multi-scale approach for pre-
tral California shelf. Estuar Coast Shelf Sci 83:629−637 dicting fish species distributions across coral reef sea-
Laidig TE, Krigsman LM, Yoklavich MM (2013) Reactions of scapes. PLoS ONE 6:e20583
fishes to two underwater survey tools, a manned sub- Pittman SJ, Costa BM, Battista TA (2009) Using lidar
mersible and a remotely operated vehicle. Fish Bull 111: bathymetry and boosted regression trees to predict the
54–67 diversity and abundance of fish and corals. J Coast Res
Levin PS, Schwing FB (eds) (2011) Technical background for (Spec Issue) 53:27−38
an integrated ecosystem assessment of the California Punt A, Ralston S (eds) (2007) A management strategy eval-
Current: groundfish, salmon, green sturgeon, and eco- uation of rebuilding revision rules for overfished rockfish
system health. US Dep Commer NOAA Tech Memo stocks. Alaska Sea Grant College Program, AK-SG-07-01
NMFS-NWFSC-109 R Development Core Team (2011), R: a language and envi-
Love MS (2006) Subsistence, commercial, and recreational ronment for statistical computing. The R Foundation for
fisheries. In: Allen LG, Pondella DJ, Horn MH (eds) The Statistical Computing, Vienna. www.r-project.org
ecology of marine fishes: California and adjacent waters. ➤ Ralston S, MacFarlane BR (2010) Population estimation of
University of California Press, Berkeley, CA, p 567−594 bocaccio (Sebastes paucispinis) based on larval produc-
Love MS (2011) Certainly more than you want to know tion. Can J Fish Aquat Sci 67:1005−1020
about the fishes of the Pacific coast. Really Big Press, Ralston S, Pearson DE, Field JC, Key M (2010) Documenta-
Santa Barbara, CA tion of the California catch reconstruction project. NOAA
Love MS, Yoklavich M (2006) Deep rock habitats. In: Allen Tech Memo NOAA-TM-NMFS-SWFSC-461
LG, Pondella DJ, Horn MH (eds) The ecology of marine ➤ Roberts JJ, Best BD, Dunn DC, Treml EA, Halpin PN (2010)
fishes: California and adjacent waters. University of Cali- Marine Geospatial Ecology Tools: an integrated frame-
fornia Press, Berkeley, CA, p 253−266 work for ecological geoprocessing with ArcGIS, Python,
Love MS, Morris P, McCrae M, Collins R (1990) Life history R, MATLAB, and C++. Environ Model Softw 25:
aspects of 19 rockfish species (Scorpaenidae: Sebastes) 1197−1207
from the Southern California Bight. NOAA Tech Rep ➤ Wilson MF, O’Connell B, Brown C, Guinan JC, Grehan AJ
NMFS-87 (2007) Multiscale terrain analysis of multibeam bathy-
Love MS, Yoklavich M, Thorsteinson L (2002) The rock- metry data for habitat mapping on the continental slope.
fishes of the Northeast Pacific. University of California Mar Geod 30:3−35
Press, Berkeley, CA ➤ Wood SN (2004) Stable and efficient multiple smoothing
➤ Love MS, Yoklavich M, Schroeder DM (2009) Demersal fish parameter estimation for generalized additive models.
assemblages in the Southern California Bight based on J Am Stat Assoc 99:673−686
visual surveys in deep water. Environ Biol Fishes 84: Yoklavich M, Starr R, Steger J, Greene HG, Schwing F, Mal-
55−68 zone C (1997) Mapping benthic habitats and ocean
➤ Lundblad ER, Wright DJ, Miller J, Larkin EM and others currents in the vicinity of central California’s Big Creek
(2006) A benthic terrain classification scheme for Ameri- Ecological Reserve. NOAA Tech Memo, NOAA-TM-
can Samoa. Mar Geod 29:89−111 NMFS-SWFSC-245
➤ Miller RR, Field JC, Santora JA, Schroeder ID and others Yoklavich MM, Greene HG, Cailliet GM, Sullivan DE, Lea
(2014) A spatially distinct history of the development of RN, Love MS (2000) Habitat associations of deep-water
California groundfish fisheries. PLoS ONE 9:e99758 rockfishes in a submarine canyon: an example of a natu-
Montgomery DC, Peck EA (1992) Introduction to linear ral refuge. Fish Bull 98:625−641
regression analysis, 2nd edn, Probability and Statistics Yoklavich MM, Cailliet GM, Lea RN, Greene HG, Starr R,
Series. John Wiley & Sons, New York, NY DeMarignac J, Field J (2002) Deepwater habitat and fish
➤ Moore CH, Harvey ES, Van Neil K (2010) The application of resources associated with the Big Creek Marine Ecolog-
predicted habitat models to investigate the spatial eco- ical Reserve. CCOFI Rep 43:120−140
logy of demersal fish assemblages. Mar Biol 157: ➤ Yoklavich MM, Love MS, Forney KA (2007) A fishery-
2717−2729 independent assessment of an overfished rockfish stock,
➤ O’Farrell MR, Yoklavich MM, Love MS (2009) Assessment cowcod (Sebastes levis), using direct observations from
of habitat and predator effects on dwarf rockfishes an occupied submersible. Can J Fish Aquat Sci 64:
(Sebastes spp) using multi model inference. Environ Biol 1795−1804
Fishes 85:239−250 ➤ Young MA, Iampietro PJ, Kvitek RG, Garza CD (2010) Mul-
➤ Parker SJ, Berkeley SA, Golden JT, Gunderson DR and oth- tivariate bathymetry-derived generalized linear model
ers (2000) Management of Pacific rockfish. Fisheries 25: accurately predicts rockfish distribution on Cordell
22−30 Bank, California, USA. Mar Ecol Prog Ser 415:247−261
PFMC (Pacific Fishery Management Council) (2011) Pacific Zuur A, Ieno EN, Walker N, Saveliev AA, Smith GM (2009)
coast groundfish fishery management plan for the Mixed effects models and extensions in ecology with R.
California, Oregon, and Washington groundfish fish- Springer, New York, NY
ery. PFMC, Portland, OR. www.pcouncil.org/groundfish/ ➤ Zuur AF, Ieno EN, Elphick C (2010) A protocol for data
fishery-management-plan/ exploration to avoid common statistical problems.
Pinheiro JC, Bates DM (2009) Mixed-effects models in S and Methods Ecol Evol 1:3−14
Editorial responsibility: Jake Rice, Submitted: February 5, 2015; Accepted: July 28, 2015
Ottawa, Ontario, Canada Proofs received from author(s): November 12, 2015