supplementary materials for - science advances · 3, and cv 2015 representing the auc calculated on...
TRANSCRIPT
advances.sciencemag.org/cgi/content/full/4/5/eaar3001/DC1
Supplementary Materials for
A dynamic ocean management tool to reduce bycatch and support
sustainable fisheries
Elliott L. Hazen, Kylie L. Scales, Sara M. Maxwell, Dana K. Briscoe, Heather Welch, Steven J. Bograd,
Helen Bailey, Scott R. Benson, Tomo Eguchi, Heidi Dewar, Suzy Kohin, Daniel P. Costa,
Larry B. Crowder, Rebecca L. Lewison
Published 30 May 2018, Sci. Adv. 4, eaar3001 (2018)
DOI: 10.1126/sciadv.aar3001
The PDF file includes:
Supplementary Methods
table S1. Species-specific model deviance explained and cross-validation using
area under the curve statistics.
fig. S1. Kernel density plot of fisheries effort and tracking data for leatherback
turtles, California sea lions, and blue sharks.
fig. S2. Sample track with three randomly selected pseudotracks for all three
satellite-tracked species.
fig. S3. Partial response curves from boosted regression trees for sea surface
temperature, bathymetry, chl-a, and SSHa across all species models.
fig. S4. Species-specific predictions with error bounds from boosted regression
tree model fitting process.
fig. S5. Time series of species habitat in a normal (2012) and anomalously warm
(2015) year.
fig. S6. Sensitivity analysis of EcoCast bycatch and integrated risk under varying
species weightings to highlight their influence on the final product.
fig. S7. Operational tool for exploring EcoCast weightings available to managers
to assess how varying scenarios change the integrated risk surface.
Other Supplementary Material for this manuscript includes the following:
(available at advances.sciencemag.org/cgi/content/full/4/5/eaar3001/DC1)
movie S1 (.mp4 format). Animation of daily bycatch predictions for the August to
December 2012 fishing season, with red pixels representing high risk and white
representing low risk.
movie S2 (.mp4 format). Animation of daily bycatch predictions for the August to
December 2015 fishing season, with red pixels representing high bycatch risk and
white representing low risk. movie S3 (.mp4 format). Animation of daily integrated predictions for the August
to December 2015 fishing season, with red pixels representing high bycatch risk
and low target catch and with blue pixels representing high target catch and low
bycatch risk.
movie S4 (.mp4 format). Animation of daily integrated predictions for the August
to December 2015 fishing season, with red pixels representing high bycatch risk
and low target catch and with blue pixels representing high target catch and low
bycatch risk.
Supplementary Methods
We explored three modeling frameworks: generalized linear mixed models (GLMMs),
generalized additive mixed models (GAMMs), and boosted regression trees (BRTs). GLMMs
allow non-linear relationships between predictor and response variables using a link function,
while GAMMs use penalized regression splines instead of a transformed linear relationship,
allowing for more flexibility in model structure. BRTs are an extension to classification and
regression trees that use boosting to optimize the partitioning of variance (34). We found that
BRTs consistently outperformed GLMMs and GAMMs, thus we chose BRTs to generate the
predictive habitat suitability surfaces underlying EcoCast (19). For the models built using
tracking datasets (blue shark, California sea lion, leatherback turtle), we iteratively and randomly
resampled without replacement from presence and absence datasets 1000 times to quantify data
variability, estimate process error, reduce serial autocorrelation and eliminate bias resulting from
arbitrarily selecting individual correlated random walks to match to tracks. For models built
using observer data, we used the presence (set with catch) and absence (set but no catch) of each
of the focal species (swordfish, blue shark) as the response. We included all of the available
environmental predictors in the BRT framework to partition the deviance accordingly rather than
parametric-based information criterion approaches (34). We observed some variability in the
contribution of environmental variables to the overall predictive capabilities of each model
(Table S1). For instance, for blue sharks, SST and bathymetry changed ranking of variable
importance between the models built using fisheries observer data and those based on tracking
data. The difference in variable importance among tag-based and observer-based models are
representing drivers of blue shark habitat use compared to fisheries effort (both species must be
present and gillnet must be set) and ultimately catchability (depth overlap between blue sharks
and gillnet) respectively. Given the model domain, the peak in leatherback habitat suitability at
SST values of 25°C is likely a reflection of migration habitat in the southwest portion of the
study areabut not necessarily overlap with the fishery. A fishery interaction model compared to
the current species distribution model would likely show a much lower peak in SST given where
interactions have been most common.
To validate our model predictions for all five models (swordfish and blue shark observer-based
models and blue shark, leatherback turtle and California sea lion tracking-based models), we
used several formulations of k-fold cross-validation, with area under the Receiver Operating
Characteristic curve (AUC) as a diagnostic (Table S1).Model AUC was calculated using a 75%
training / 25% test approach to estimate the sensitivity and specificity of our predictions (19,34).
In addition, a leave-one-out cross-validation approach was taken to validate the predictive
capabilities of each model over each of the years for which we had data. In each iteration, we left
out a single year from the model fitting process and calculated a mean AUC from each retained
year. AUC is not a perfect evaluation metric, thus we ensured that ecological realism was also
considered in evaluating models.Given the extreme conditions and low effort of 2015, observer
data from 2015 were not including in the model-fitting and instead were treated as an additional
independent dataset for cross-validation.AUC statistics were highest for the leatherback and sea
lion tracking models, and the blue shark and swordfish models performed much better than
random. In addition, the models were more successful in predicting missing data across years
than when a specific year was left out, however the blue shark observer model performed better
on 2015 than when predicting across modeled years and via random sampling. This approach
also ensured that the observer models were able to temporally extrapolate to novel conditions
providing confidence in its use as an operational tool.We created daily predictions and 95%
confidence intervals by speciesto highlight areas where predicted distributions and uncertainty
were highest (Figure S2). Confidence intervals were calculated across 10 boosted regression tree
model fits to account for model stochasticity (19,34).
Operationalization
To allow for applied use by managers and fishers, we operationalized our dynamic ocean
management approach into a real-time tool that produces daily predictions. As management
priorities can shift based on new stock data or high levels of mortality in a species of concern,
the EcoCast tool allows adjustment of weightings in reflection of new information on bycatch
risk in the middle of a fishing season (Figure S7). The tool also allows for exploration of
historical predictions, has the ability to zoom, and can show additional datasets such as nautical
charts (https://heatherwelch.shinyapps.io/ecocastapp/). The EcoCast product is hosted online and
updated daily (http://oceanview.pfeg.noaa.gov/ecocast/) and thus can be delivered directly to
fishers when minimal internet is available. In addition, when risk is highest, managers can close
the areas of greatest concern to reduce risk yet still considering areas of high swordfish catch.
table S1. Species-specific model deviance explained and cross-validation using area under
the curve statistics. We calculated AUC using 75% of the data as training with 25% as testing in
column 2, CV by holding out each year as test and using the remaining years as train in column
3, and CV 2015 representing the AUC calculated on an entirely novel year.
Table S1.
Dev. AUC CV CV 2015
swordfish 12.7 0.72 0.66 0.69
blue shark 12.2 0.73 0.67 0.82
blue shark tracking 16.3 0.76 0.73 NA
leatherback 49.7 0.93 0.85 NA
sea lion 69.7 0.86 0.81 NA
fig. S1. Kernel density plot of fisheries effort and tracking data for leatherback turtles,
California sea lions, and blue sharks. Map of tracking and DGN fisheries observer data as
kernel densities. The Pacific Leatherback Conservation Area (PLCA) is shown in grey dashed
lines. Fisheries observer data are shown as 50 and 95% kernel densities from low effort (yellow)
to high (red). Home ranges for the three tagged species are shown as dashed lines with
leatherback turtles (green - 50% kernel density), blue sharks (blue - 50% kernel density), and
California sea lions (brown – 50 and 95% kernel densities).
fig. S2. Sample track with three randomly selected pseudotracks for all three satellite-
tracked species. Randomly selected example track line for a single sea lion (brown), leatherback
turtle (green), and blue shark (blue) with 3 example pseudotracks shown as solid points.
−150 −140 −130 −120 −110
20
25
30
35
40
45
50
fig. S3. Partial response curves from boosted regression trees for sea surface temperature,
bathymetry, chl-a, and SSHa across all species models. Response curves from Boosted
Regression Trees for A) Sea Surface Temperature (SST) in °C, B) Bathymetry in m, C)
Chlorophyll-a in mg/m3, and D) Sea Surface Height anomaly (SSHa) in cm. The five curves
represent swordfish (black), sea lions (brown), leatherbacks (green), blue shark tracking (dark
blue) and blue shark observer data (light blue).
fig. S4. Species-specific predictions with error bounds from boosted regression tree model
fitting process. Boosted regression tree model predictions including 95% confidence intervals as
upper and lower bounds around the mean for A-C) swordfish observer data, D-F) leatherback
turtle tracking data, G-I) California sea lion tracking data, J-L) blue shark tracking data, and M-
O) blue shark observer data. Dark blue represents higher predicted species presence while light
blue represents lower predicted presence.
fig. S5. Time series of species habitat in a normal (2012) and anomalously warm (2015)
year. Species habitat time series (4 panel) - Time series of top quartile of predicted habitat as
percentage of the total study area (-115° to -130° E longitude and 30° to 50° N latitude) were
calculated to examine whether finer-scale temporal closures could be more successful than
current PLCA closure dates. (A) Sea lions (brown) and (B) blue sharks (light and dark blue)
were predicted to have the smallest habitat range, followed by (C) leatherback sea turtles (green),
with (D) swordfish (black) the most broadly distributed. All individual bycatch time series
showed a stable or decreasing trend in the top quartile of habitat throughout the fishing season
except for sea lions in 2015. This is in contrast to swordfish that showed a stable trend in 2012
and an increasing trend for 2015 throughout the season.
fig. S6. Sensitivity analysis of EcoCast bycatch and integrated risk under varying species
weightings to highlight their influence on the final product. EcoCast sensitivity analysis
showing how integrated results vary based on species weightings for bycatch only models (A,D)
and integrated EcoCast models (B,C,E,F). Areas of high bycatch risk are shown in red with low
risk in white, while integrated EcoCast risk varies from low target catch / high bycatch in red to
high target catch / low bycatch in blue. Transparency of species’ silhouettes indicate the relative
influence in the surface with less transparent indicating less overall influence in the integrated
surface.
fig. S7. Operational tool for exploring EcoCast weightings available to managers to assess
how varying scenarios change the integrated risk surface. Tool for serving updated EcoCast
product with manager-adjustable risk weightings. This tool also allows viewing of previous
predictions, adding addition data layers, and downloading of raw data.