Effect of a Mammal Excluder Device on Trawl Catches of Salmon and Other Pelagic Animals

Abstract

Concern about bycatch of marine mammals by fishery research gear has led to the use of mammal excluder devices (MEDs) in some surveys. We measured the effect of a rigid‐grate MED on trawl fishing effectiveness for juvenile Pacific salmon Oncorhynchus spp. and other upper‐pelagic species in the northern California Current. Catches of 11 taxa were compared among sets of a pelagic rope trawl without an MED and with an MED in two orientations: angled upward or angled downward. We estimated differences in overall catch ratios and size selectivity for each MED orientation compared to the unmodified net. The MED orientation had a species‐specific influence on both catch ratio and size selectivity, with median ratios ranging from 0.0 to 1.8 for the upward orientation and from 0.4 to 2.7 for the downward orientation. Video observations helped to explain performance differences and may help to improve MED design. Use of an MED in trawl surveys may substantially increase or decrease the catch of various taxa, contributing further uncertainty to the analysis of survey time series for either abundance or catch composition.

Conflicts often arise in species conservation when actions benefiting one species may harm another. This type of conflict is perhaps epitomized by the interactions between marine mammals and Pacific salmon Oncorhynchus spp. along the U.S. West Coast. Several species of marine mammals occur along the coast; all are protected under the U.S. Marine Mammal Protection Act, with eight species also listed under the U.S. Endangered Species Act (ESA; NMFS ). Additionally, several West Coast distinct populations of Pacific salmon are protected under the ESA (NMFS ), and salmon and other harvested species are regulated under the Magnuson–Stevens Fishery Conservation and Management Act. The National Marine Fisheries Service (NMFS) administers protections under all three laws. The most apparent mammal–salmon conflicts arise from predator–prey interactions, with examples both of locally high mammal abundance limiting the survival of salmon (Adams et al. ; Chasco et al. ) and low salmon abundance limiting the growth and reproduction of mammals (Vélez‐Espino et al. ). This study was developed in response to a subtler interaction: the potential effects of salmon conservation research activities on marine mammals via bycatch in research survey trawls.

The NMFS Northwest Fisheries Science Center and Southwest Fisheries Science Center jointly conduct studies of juvenile salmon and other small pelagic fishes along the West Coast from California to Southeast Alaska. These studies aim to improve our understanding of factors affecting growth and survival in support of salmon recovery planning and harvest management (Fisher et al. ). Increasing concerns about marine mammal bycatch in surface trawls led to the design of a mammal excluder device (MED; Dotson et al. ) and its required use during all research surveys. Use of the MED was mandated by agency policy, with little evaluation of either its effectiveness at reducing mammal bycatch or its effects on target species catches. This study is directed at providing information for the latter issue.

To evaluate the effect of MEDs on survey data time series, we conducted gear comparison trials with and without the MED installed in the pelagic rope trawls used during surveys. Gear comparisons were conducted in the northern California Current during six cruises between 2009 and 2015. Dotson et al. () described the MED and analyzed its use during the 2009 sampling. They concluded that the device had no discernible effect on size distributions of the main species caught (Pacific Sardine Sardinops sagax, Jack Mackerel Trachurus symmetricus, Northern Anchovy Engraulis mordax, Pacific Herring Clupea pallasii, and Humboldt squid Dosidicus gigas) and no significant effect on total catch, although the small sample size resulted in little power to detect differences.

Here, we present an analysis of both overall catch rate and size differences between pelagic rope trawls based on data from trials in 2011, 2014, and 2015 using a trawl with and without an MED installed. The objectives and sampling protocols were somewhat different in the 2009 studies, so we have not included the 2009 data. The 2009 study was focused on sampling Pacific Sardine, with an emphasis on species composition and size distributions rather than on estimating abundance (Dotson et al. ).

The purpose of the 2011–2015 trials was specifically to estimate correction factors for overall catch rates of pre‐defined age‐classes of juvenile salmon so that abundance data from surveys using the MED could be compared with those using the unmodified net. The NMFS has used the same trawl design for regular surveys conducted for nearly 20 years to produce a time series of juvenile salmon abundance indices (Peterson et al. ). Thus, any changes in catchability related to the MED would limit our ability to interpret the time series. This is a widespread issue for any change to gear or vessel used in fish surveys (Pelletier ).

Methods

Field sampling

Sampling was conducted using a Nordic 264 rope trawl (Nor'Eastern Trawl Systems, Bainbridge Island, Washington). The net is widely used in studies of juvenile forage fish and juvenile salmon in the northern California Current and coastal waters of Alaska and British Columbia. When deployed, the trawl has an effective opening of approximately 20 m high × 30 m wide and a length of about 200 m.

The MED consists of an aluminum rectangular grate that is 155 cm long and 112 cm wide, with 13‐cm spacing between vertical bars. The grate is installed in a 10‐m‐long × 112‐cm square net section (100‐mm mesh consisting of 5‐mm polyethylene twine) sewn into the trawl just ahead of the cod end. As originally designed, it is positioned sloping upwards at a 46–47° angle leading to an escape opening covered by a flap of 38‐mm web; the flap is designed to open when something large pushes against it. The intention of this design is to pass small fishes through the grate to the cod end, while allowing larger animals, such as marine mammals and sharks, to escape through the opening. Full details of the design are provided by Dotson et al. ().

Trials with this configuration conducted in 2011 and 2014 suggested a significant loss of some target species when using the MED (Weitkamp ). Video observations during these trials suggested that a downward‐sloping grate with an escape opening at the bottom of the net might provide better retention of target species. For this reason, in 2015 trials, the MED was configured in a downward orientation, with the entire MED net section flipped upside‐down. Foam floats were added to the cover flap to hold it up against the bottom of the trawl.

General field protocols followed the standard methods for coastal salmon surveys conducted by the Northwest Fisheries Science Center (Fisher et al. ; Peterson et al. ). For each tow, the start and end locations and times (from the end of deployment to the start of retrieval) were recorded, all captured animals were identified and enumerated, and a subsample of each species was measured. We conducted a total of 86 tows off the southern Washington and northern Oregon coast over a total of 11 d during May and July 2011, June and July 2014, and July 2015. The net was towed in daylight near the surface (upper 20 m) for 15–30 min at speeds of 1.4–2.0 m/s (2.7–3.8 knots) over distances ranging between 1.1 and 4.9 km. Sampling sites (Figure ) were chosen based on an expected high abundance of juvenile salmon from past sampling experience (Peterson et al. ) and juvenile salmon abundance patterns observed during coastal surveys conducted 1–2 weeks prior to our test cruises (C. A. Morgan, Oregon State University, Corvallis, personal communication, 2011–2015). During each cruise, gear types were alternated: one or two samples were taken with the MED installed in the net, followed by one or two samples taken without the MED. This pattern was continued through each sampling day. This protocol was a departure from the initial experimental design, which called for paired side‐by‐side tows. Budget and scheduling constraints prevented contracting with a second vessel, so we compromised on the alternating‐gear, unpaired sampling design. Individual size was measured for all species caught. Size subsampling rates varied by species. For targeted salmon, all fish were measured when catches were less than 101 individuals per species; when catches exceeded 101 individuals per species, a random subsample of 100 individuals was measured. For nontarget species, subsampling rates varied depending on the numbers caught and time available; for large catches, between 10 and 100 individuals of each nontarget species were measured. We measured the FL of fish to the nearest 1 mm, the bell diameter of jellies to the nearest 5 mm, and the mantle length of squid to the nearest 1 mm.

Map of sampling locations off the coast of southern Washington and northern Oregon. Symbols identify survey cruises (circles = May 2011; squares = July 2011; diamonds = June–July 2014; triangles = July 2015). Boxes with letters (A–E) identify sampling locations used in the statistical analysis (see Table ). Depth contours are shown at 50 and 100 m.

Sampling was conducted at one or two sites each day, and equal numbers of samples were taken with each gear at each site. Of the 86 tows completed, 43 were conducted with the unmodified net, 27 were performed with the MED angled upward (2011 and 2014), and 16 were conducted with the MED angled downward (2015). Key data for all hauls are provided in Table S1 available separately online.

Species selection

We analyzed catch rate data (number caught per kilometer trawled) and size distributions for all taxa for which at least 100 individuals were caught (total across all samples) and for which more than one individual was caught in at least three of the four cruises. Because our surveys were focused on juvenile Coho Salmon O. kisutch and Chinook Salmon O. tshawytscha, in defining “taxa” for analysis, we subdivided those two species into age‐groups: subyearling and yearling first‐ocean‐year juveniles; and “subadults” that have spent one or more winters in the ocean. Age‐groups were determined by ranges in measured FL using criteria defined by Peterson et al. () and Teel et al. (). With these definitions, 11 taxa/age‐groups met the minimum sampling criteria (Table S2): Chinook Salmon (subyearling, yearling, and subadult), Coho Salmon (yearling and subadult), Chum Salmon O. keta, Northern Anchovy, Pacific Herring, market squid Doryteuthis opalescens, Pacific sea nettle Chrysaora fuscescens, and water jellies Aequorea spp. Total number caught, total number measured, and average subsampling rate for these species are presented in Table .

Number caught, number measured, and average subsampling rate (SSR) for selected species or age‐groups by gear type (mammal excluder device [MED] with downward escape, MED with upward escape, or standard net [no MED])

Taxon	Number caught	Number measured	Average SSR
Downward MED	Upward MED	No MED	Downward MED	Upward MED	No MED	Downward MED	Upward MED	No MED
Chinook Salmon subyearling	191	713	1,091	142	713	1,035	0.74	1.00	0.95
Chinook Salmon yearling	90	223	365	90	223	363	1.00	1.00	1.00
Chinook Salmon subadult	28	36	59	28	36	59	1.00	1.00	1.00
Chum Salmon	54	0	173	54	0	173	1.00	NA	1.00
Coho Salmon yearling	38	24	194	38	24	194	1.00	1.00	1.00
Coho Salmon subadult	41	25	61	41	25	61	1.00	1.00	1.00
Northern Anchovy	2	200	3,073	2	152	241	1.00	0.76	0.07
Pacific Herring	1,755	58	563	41	31	73	0.02	0.53	0.13
Water jellies	53,291	528	71,472	385	306	639	0.01	0.58	0.01
Market squid	2,160	258	8,285	347	189	582	0.16	0.73	0.07
Pacific sea nettle	21	1,285	1,422	21	536	623	1.00	0.42	0.44

Statistical methods

We assessed two potential species‐specific effects of the MED: overall catch rate (gear efficiency) and size selectivity. All computations were conducted using scripts written in the R language (R Foundation for Statistical Computing, Vienna). Detailed statistical methods, data, and R scripts for the analyses are available in a GitHub repository (https://github.com/tcwain/GearComparisonAnalysis2018).

To compare overall catch rates of the net with versus without the MED, we estimated the catch ratio (A), defined as the ratio of CPUE expected with the MED installed to the CPUE expected with the standard net (STD; without the MED):[Image Omitted. See PDF] where E(·) represents the statistical expected value. For this application, CPUE is measured as the number caught per distance towed, expressed as individuals per kilometer.

Because aquatic species generally occur in patches at varying spatial scales, catch distributions are often highly skewed, with frequent values of zero. To cover a range of possible assumptions about spatial–temporal patterns of the catch, we considered a number of approaches. These included both paired‐sample and blocked statistical designs and both parametric and nonparametric methods. After examining aggregation patterns in fish catch data and performing some preliminary comparisons of methods, we selected a general linear model (GLM) approach using a blocked sampling design and a negative binomial error distribution.

General linear models allow fitting linear or quasi‐linear statistical models to data with non‐Gaussian errors (McCullagh and Nelder ) and are particularly useful for ecological count data, which may have error structures close to Poisson or negative binomial distributions (Elliott ). For analyzing fish survey data, GLMs have the advantage of flexibly fitting regression or ANOVA models by using catch data directly (without transformations) and can easily account for variation in effort or other confounding factors by using offsets within the model structure (Lewy et al. ; Heino et al. ).

We developed a GLM analysis of deviance (ANODEV) model for contrasting gear effects nested within sampling blocks defined by date and site (Table ). This approach explicitly incorporates patchiness of the catch in two ways. First, at broad spatial–temporal scales, the block structure accounts for variation in fish community composition among sites and across time. Second, at smaller scales, the negative binomial error distribution accounts for aggregations due to schooling or other behaviors.

Data blocks defined for the general linear model–analysis of deviance model. Locations A–E are identified in Figure 

Block	Date	Location	Number of tows
1	May 19, 2011	A	10
2	May 20, 2011	A	10
3	Jul 27, 2011	B	4
4	Jul 27, 2011	A	4
5	Jul 28, 2011	A	10
6	Jun 29, 2014	B	4
7	Jun 30, 2014	C	4
8	Jul 1, 2014	C	8
9	Jul 4, 2015	D	8
10	Jul 5, 2015	D	8
11	Jul 6, 2015	E	8
12	Jul 7, 2015	D	8

The model links observed numbers caught in a single tow to gear type (a three‐level factor: standard gear, MED oriented upward, or MED oriented downward) and sampling block (a factor indicating date and location), with an offset correcting for variations in distance towed:[Image Omitted. See PDF]

The model was fitted using negative binomial errors and a logarithmic link function, which is the natural link for the negative binomial distribution (McCullagh and Nelder ). We estimated the negative binomial shape parameter (θ) simultaneously with the catch ratio statistics using the function “glm.nb” in the R package “MASS” (Venables and Ripley ). Ideally, this parameter would have been estimated a priori for our target species by using repeated samples with the standard gear, but we lacked sufficient repeat‐sample data to do that reliably. Errors in estimating the shape parameter are not included in the GLM ANODEV statistics reported here; for that reason, our reported confidence intervals (CIs) may be somewhat narrow.

The ANODEV model provides a measure of the catch ratio as the estimated logarithmic‐scale coefficient for Gear. To estimate the catch ratio A, the mean estimate is back‐transformed with a correction for estimation error:[Image Omitted. See PDF] where μ and σ are the mean and SD of the estimated Gear coefficient. Medians and CIs were estimated by back‐transforming quantiles of the Student's t distribution for the estimated coefficient.

For size‐selectivity comparisons, size frequencies were summed in 5‐mm bins, and the resulting frequencies were plotted for visual inspection. As an overall test for gear differences, we applied a Kolmogorov–Smirnov (KS) two‐sample test (Sprent ) comparing raw number‐at‐size distributions for samples with and without an MED. We also fitted size‐selectivity curves to the data.

Quantitative estimation of size‐selectivity curves for fishing gear has a long history but has recently evolved rapidly from traditional parametric fits of logistic or probit curves to incorporate nonparametric fitting and bootstrapping for CIs (Millar ) and the use of more flexible parametric polynomial or nonparametric smooth selectivity curves (Holst and Revill ; Krag et al. ; Herrmann et al. ; Kotwicki et al. ).

We used a third‐order polynomial regression model to describe the relationship of gear selectivity to size. This model is similar to those used by Krag et al. () and Kotwicki et al. (). We tried other models, including fourth‐order polynomials and nonparametric smooth‐curve generalized additive models, but we found that the added complexity was not supported by our data.

Because our samples were unpaired and numbers caught in individual hauls were often low, we fitted the model to total CPUE at size across all hauls for each gear. To fit the model for each species, the catch at size within each haul (C_h,s) was expanded by the haul‐specific subsampling ratio (ssr_h). For each gear (g), CPUE by size was then computed as the sum of adjusted counts ($C_{h,s}^{\ast }$) divided by the sum of effort for each haul:[Image Omitted. See PDF][Image Omitted. See PDF] where the sums are over all hauls with a given gear type. We then computed the overall catch comparison rate for two gears (r_s; Krag et al. ) as[Image Omitted. See PDF] where gear 1 is assumed to be the unmodified net, and gear 2 is the net with the MED installed.

We then fitted a third‐order polynomial model to logit‐transformed catch comparison rates,[Image Omitted. See PDF] where ε represents Gaussian error. The model was fitted using the R function “glm” for the binomial family with logit links. For comparison with the bulk catch ratio estimates, we converted the fitted catch comparison rates to the size‐specific catch ratio (CR_s) as in Kotwicki et al. ():[Image Omitted. See PDF]

We used a double‐bootstrap procedure (Millar ) to compute approximate CIs around the estimated curves. In this procedure, each bootstrap replicate includes resampling of both between‐haul and within‐haul variation. To represent between‐haul variation, individual hauls equal in number to the original data were chosen with replacement from among all hauls. To represent within‐haul variation, within each chosen haul, all individuals caught were resampled with replacement, and the resulting size‐frequency distributions and CPUEs were recomputed. The regression model was refitted to this resampled data set, and the predicted size‐specific catch ratio was computed. One‐thousand bootstrap replicates were used to generate approximate centered 90% CIs on predicted catch ratios via the percentile method (Efron :chapter 10).

Video camera methods

In 2014 and 2015, we mounted video cameras (GoPro Hero 3+  or 4+  models; GoPro, Inc., San Mateo, California) on the net to monitor fish behavior around the MED. Camera locations varied over the course of the cruises, but at a minimum, one camera was mounted inside the throat of the net, facing forward to observe fish approaching the cod end (Figure , A). For tows with the MED installed, one camera was mounted outside of the net behind the MED, facing forward to observe the MED escape opening (Figure , D). Because trawling was conducted near the surface during daylight, we relied on ambient light.

Placement of cameras in the mammal excluder device (MED) net section. Three cameras (A, B, and C) were attached inside the net mesh, with the field of view (dashed lines) covering the approach to the MED (A and B) and the MED grate (C). The exit camera (D) was attached to the outside of the top mesh behind the fine‐mesh escape flap, with the field of view covering the escape opening.

The resulting videos were used for general observations of gear performance and fish behavior. We did not try to classify or quantify these behaviors. We did try to count animals entering the cod end and those exiting from the escape opening, but video quality was not sufficient for reliable identification of all fish to the species level, and variability in turbidity, ambient light, and fish behavior made reliable counts impossible in most tows, even for broad taxonomic groups; in only a few tows were we able to count juvenile salmon both entering the cod end and exiting via the escape opening. Because the video observations are general and somewhat anecdotal, we do not include them in the Results section, but we do include some observations and speculations about fish behavior in the Discussion.

Results

Overall Catch Ratios

Catch ratios varied widely among species and MED orientations (Table ; Figure ). With the MED upward configuration, 5 of 11 taxa had median ratios near 1.0 (between 0.75 and 1.25), 5 taxa had ratios substantially below 1.0 (reduced catch with the MED), and 1 taxon (water jellies) had a ratio above 1.0 (enhanced catch with the MED). In some cases, the reduction in catch was extreme: Chum Salmon, Northern Anchovy, and Pacific Herring all exhibited greater than 90% reduction in catch with the MED, and yearling Coho Salmon had a reduction of about two‐thirds.

Estimated catch ratio (with 90% confidence interval [CI]) and negative binomial shape parameter (θ; with SE) for the mammal excluder device (MED) in an upward or downward orientation (NA = no estimate is available)

Taxon	θ (SE)	MED orientation	Median catch ratio	Mean catch ratio	90% CI
Chinook Salmon subyearling	1.3 (0.3)	Upward	0.85	0.89	0.54–1.37
Downward	0.49	0.51	0.30–0.80
Chinook Salmon yearling	1.5 (0.3)	Upward	0.76	0.79	0.47–1.21
Downward	1.18	1.24	0.71–1.98
Chinook Salmon subadult	11.5 (15.3)	Upward	1.03	1.07	0.65–1.65
Downward	1.36	1.45	0.77–2.43
Coho Salmon yearling	1.5 (0.4)	Upward	0.32	0.34	0.17–0.59
Downward	0.67	0.72	0.35–1.27
Coho Salmon subadult	3.7 (2.1)	Upward	0.63	0.66	0.38–1.05
Downward	2.69	2.87	1.48–4.89
Chum Salmon	1.3 (0.4)	Upward	0.00	0.00	NA
Downward	0.37	0.39	0.24–0.59
Northern Anchovy	0.3 (0.1)	Upward	0.08	0.10	0.03–0.21
Downward	2.30	3.89	0.42–12.69
Pacific Herring	0.1 (NA)	Upward	0.05	0.07	0.01–0.20
Downward	2.38	2.99	0.77–7.33
Market squid	0.9 (0.2)	Upward	0.82	0.88	0.45–1.49
Downward	0.44	0.48	0.22–0.89
Pacific sea nettle	2.2 (0.5)	Upward	0.91	0.94	0.61–1.35
Downward	0.59	0.65	0.29–1.21
Water jellies	0.9 (0.1)	Upward	1.79	1.92	0.96–3.35
Downward	0.64	0.69	0.34–1.19

Estimated catch ratios for mammal excluder devices (MEDs) oriented upward (red, open boxes) and downward (blue, shaded boxes). The horizontal bar marks a ratio of 1.0, where catch with the MED equals catch without the MED. For each species, the plus symbol represents the mean estimate, the diamond symbol denotes the median, boxes span the quartiles, and whiskers span the 0.05 and 0.95 quantiles (90% confidence interval).

With the MED downward configuration, one taxon (yearling Chinook Salmon) had a ratio near 1.0, six taxa had ratios below 1.0, and four taxa had ratios above 1.0. However, none of the species exhibited the extreme reductions seen with the upward orientation. Compared with the upward orientation, the downward orientation improved catch for seven of the taxa, but reduced catch for the other four (subyearling Chinook Salmon, market squid, Pacific sea nettle, and water jellies). For the four taxa exhibiting extreme reductions with the upward orientation, using the downward orientation either reduced the negative effect (Chum Salmon and yearling Coho Salmon) or resulted in an opposite positive effect (Northern Anchovy and Pacific Herring).

Size Selectivity

We found significant overall differences in size frequencies between the standard and MED nets for four species (KS test; Table ): Chinook Salmon, Coho Salmon, Northern Anchovy, and market squid. The fitted size‐selectivity curves (Table ; Figures , ) are more informative. With few exceptions, the KS test for overall differences and the regression model fits were consistent; when the KS results were significant, so were some of the regression parameters. Chinook Salmon were an exception, but in different directions for the two MED orientations. In three cases (Chinook Salmon, downward MED; Northern Anchovy, upward MED; and market squid, downward MED), the KS test was significant, while the regression had no significant terms; examination of the data suggested that these cases may have resulted from the KS test being very sensitive to differences in the extreme tails of the distributions. In one case (Chinook Salmon, upward MED), the opposite occurred: the KS test was not significant, while the regression had two significant terms, although the resulting curve was fairly flat and the prediction CI never crossed the axis. Coho Salmon exhibited significant size patterns for both tests conducted for each MED orientation. With the upward MED, there was a significant loss of small (100–200‐mm) fish, consistent with the differences in overall catch ratio (previous section) between yearling and subadult age‐groups. With the downward MED, there was a significant gain in catch of larger (400–600 mm) fish, also consistent with the overall catch ratio estimates for subadult Coho Salmon. The size‐selectivity curves illustrated significant catch reductions over a portion of the size range for both Chum Salmon and market squid with the MED oriented downward (Figure ), consistent with the overall catch ratio estimates.

Probabilities (P‐values) and significance levels for the analysis of size data from the upward‐ or downward‐oriented mammal excluder device (MED): Kolmogorov–Smirnov (KS) two‐sample tests for differences in size‐frequency distributions and parameters of the third‐order polynomial regression fits (*P < 0.05; **P < 0.01; ***P < 0.001; nd = nonsufficient data)

Taxon	Upward MED	Downward MED
KS test	Model parameters	KS test	Model parameters
Chinook Salmon	0.776	β0: 0.011* β1: 0.014* β2: 0.067 β3: 0.184	0.000***	β0: 0.110 β1: 0.074 β2: 0.078 β3: 0.084
Chum Salmon	nd	nd	0.407	β0: 0.820 β1: 0.779 β2: 0.742 β3: 0.707
Coho Salmon	0.000***	β0: 0.000*** β1: 0.001*** β2: 0.002** β3: 0.004**	0.001**	β0: 0.297 β1: 0.180 β2: 0.086 β3: 0.049*
Northern Anchovy	0.000***	β0: 0.463 β1: 0.500 β2: 0.540 β3: 0.577	nd	nd
Market squid	0.000***	β0: 0.006** β1: 0.013* β2: 0.033* β3: 0.065	0.009**	β0: 0.258 β1: 0.081 β2: 0.123 β3: 0.207
Pacific sea nettle	0.192	β0: 0.295 β1: 0.437 β2: 0.848 β3: 0.596	nd	nd
Water jellies	0.157	β0: 0.055 β1: 0.051 β2: 0.057 β3: 0.064	0.999	β0: 0.075 β1: 0.088 β2: 0.120 β3: 0.161

Size‐frequency comparison for selected species when the upward‐oriented mammal excluder device (MED) was used. Histograms (left y‐axis) show CPUE (catch/km) in 5‐mm bins comparing the standard net (black bars below the horizontal axis at CPUE = 0) versus the upward‐oriented MED (red bars above the horizontal axis at CPUE = 0). Curves (right y‐axis) show the estimated catch ratio (black) with the bootstrap sample median (solid gray) and approximate 90% confidence interval (dashed gray).

Size‐frequency comparison for selected species when the downward‐oriented mammal excluder device (MED) was used. Histograms (left y‐axis) show CPUE (catch/km) in 5‐mm bins comparing the standard net (black bars below the horizontal axis at CPUE = 0) versus the downward‐oriented MED (blue bars above the horizontal axis at CPUE = 0). Curves (right y‐axis) show the estimated catch ratio (black) with the bootstrap sample median (solid gray) and approximate 90% confidence interval (dashed gray).

Discussion

Any change in gear or survey operations can be expected to compromise CPUE time series, and some survey programs regularly perform gear calibration studies to estimate correction factors (Pelletier ). Our study provides catch ratio estimates for a few species, which could be used to correct their data series for the effect of the MED. However, the wide CIs on catch ratio estimates will substantially reduce the precision of corrected CPUE estimates made with the MED installed. For other species, we had insufficient data to generate reliable estimates, so no corrections are possible. Additional gear comparison sampling could improve the precision of corrections, but the high level of patchiness for some of the species will limit the degree of precision that is possible (see “Statistical Considerations” below).

Loss of target species due to the MED should be balanced against the device's benefit of reducing harm to marine mammals. At present, we cannot assess that benefit because there have been no studies of the effectiveness of this device for reducing harm to marine mammals. Although the MED design (Dotson et al. ) was based on similar devices for which effectiveness has been tested (e.g., Gibson and Isaksson ; Zollett ), it is certainly not 100% effective at preventing the bycatch of marine mammals. Gibson and Isaksson () conducted flume tests of a similar MED using neoprene dummy seals and found that the device was about 50% effective at expelling dummies that hit the MED grate. Those authors considered the problems with ejection to be mostly related to the design of the dummies, and they expected that effectiveness would be higher for real mammals. In sea trials with a similar device, short‐beaked common dolphins Delphinus delphis became entangled in the MED and drowned (Northridge ). Zollett () reviewed a number of design considerations for MEDs, noting that some cetaceans (e.g., short‐beaked common dolphins) show a preference for a downward‐oriented escape opening. Catches of marine mammals during our trawl surveys are rare, but among those observed, several of the animals were entangled in the large‐mesh forward sections of the net, far ahead of where they would have benefited from the MED.

Fish Behavior

The video cameras allowed us to make some observations pertinent to the operation of the MED and to fish behavior in the net both ahead of and behind the MED. First, we observed that the escape flap was often open during tow operations, allowing fish to escape (Video S1 available separately online). The MED was designed with the expectation that flow across the net would keep the escape flap closed except when an animal pushed against it. Under our initial operational protocols, video observations showed that the flap was most often open, well away from the net. This allowed large numbers of fish to escape the net, regardless of size. For the few tows in which we were able to count salmon leaving via the escape opening, the numbers observed escaping were sufficient to explain the reduced catch with the MED installed.

In 2014, we tried a number of design changes to alleviate this problem (Weitkamp ). We settled on inverting the MED so that it angled downward, and we added flotation to the escape flap to hold it closed against the trawl. This improved performance but still did not always keep the flap closed.

We also noted a number of behaviors that may help to explain MED effects on catch for various species:

• In the trawl midsection, some species (jellies and squid) drifted passively, while most fish were actively swimming. Some species were strongly oriented facing either forward (salmon) or aftward (Jack Mackerel and others) in the net, while other species (Pacific Herring, Northern Anchovy, and Whitebait Smelt Allosmerus elongatus) tended to dart in several directions.
• Salmon are strong swimmers, capable of swimming speeds in excess of the net tow speed. They were frequently seen drifting tail‐first past the forward camera, then swimming forward past it, thus passing the camera multiple times (Video S2). Subadult and adult salmon in particular are capable of swimming out of the net mouth and were caught in relatively small numbers by the unmodified net. (This was by design: our research targets juvenile salmon, and we wished to minimize research impacts on maturing adults.)
• Reaction to the MED varied by species. Salmon approached the MED grate tail‐first and did not react to it unless they actually struck the bars. Those that passed through the grate toward the cod end remained oriented forward, with larger adult and subadult salmon appearing reluctant to pass forward through the grate to escape; they maintained a position looking forward just behind the grate, even as the net slowed during haulback (Coho Salmon in Video S3). Thus, fish that ordinarily would escape the net were caught. Other species approached the MED head‐first and reacted strongly to avoid it some distance before making contact (e.g., Jack Mackerel in Video S4). Other small fish gathered ahead of the grate and may have escaped through the side mesh of the net (Weitkamp ; J. Orsi, NMFS, personal communication).

Orientation of the MED affected catch differently for different species and age‐groups. For salmon, switching from the upward to the downward orientation increased the catch of subadult and yearling Chinook Salmon, yearling and subadult Coho Salmon, and Chum Salmon but decreased the catch of subyearling Chinook Salmon. For yearling Chinook Salmon and subadult Coho Salmon, the catch ratio switched direction, with the downward MED catching higher numbers than the standard net. Among other species, there appeared to have been a very strong effect on catch ratios for Pacific Herring and Northern Anchovy, with the upward orientation having ratios near zero and the downward orientation having ratios above 1.0. However, the high aggregation and small sample sizes for these two species resulted in wide CIs (Figure ). As might be expected for invertebrates with less swimming ability than fish, the effects of the MED on market squid and jellies tended to be small; for all three invertebrate taxa, the downward orientation produced lower median catch ratios than the upward orientation.

Enhanced catch with the MED was unexpected; we expected that adding an escape hatch would allow fish to escape, thus reducing retention, and this was the most common effect that we observed. With the downward orientation, there were three notable exceptions (subadult Coho Salmon, Northern Anchovy, and Pacific Herring). Video observations of fish behavior near the MED suggest two possible explanations for these exceptions.

First, subadult and adult salmon are capable of swimming forward out of the mouth of the net. However, as noted above, those that passed tail‐first through the grate into the cod end appeared reluctant to pass forward through the grate to escape. Thus, fish that ordinarily would escape the net may be caught in a “jail” behind the grate.

Second, for Northern Anchovy and Pacific Herring, we speculate that there is a completely different behavior at play. Both of these species form large, dense schools, and when the net encounters a school, it may entrain only a portion of it. Schooling fishes were sometimes observed outside the net schooling with those inside the net (e.g., Northern Anchovy in Video S5). It is possible that the higher densities of fish in the net attract those outside. In one instance, we observed Pacific Herring outside the net entering through the escape opening to join their schoolmates inside. If this is a common behavior, it might explain the enhanced catch with the downward MED. With the upward MED configuration, the escape opening is at or near the surface, offering less opportunity for fish to enter or re‐enter the net.

Statistical Considerations

One of the chief problems with biological sampling is identifying a reasonable statistical distribution for the observations. Classical approaches for analyzing count data assume that individuals are completely random with respect to the sampling frame—a condition known as “complete spatial randomness” (CSR). Under CSR, counts in individual trawl samples would follow a Poisson distribution (Elliott ; Diggle ).

A preliminary analysis of data from repeated trawls with this net showed that variances for most species were much higher than what would be expected for Poisson sampling, indicating that there may be substantial fish aggregation at the scale of replicate tows at the same station (T. C. Wainwright, unpublished data). In fisheries analysis, the negative binomial distribution is a common alternative for aggregate gear calibrations (Pelletier ; Wilderbuer et al. ; Lewy et al. ) and has a long history of use in analyzing ecological count data (Fisher et al. ; Anscombe ; Moyle and Lound ; Elliott ; White and Bennetts ).

For data with such distributions, GLMs are a common analytical method. Lewy et al. () and Heino et al. () both recommended GLM approaches for spatially aggregated data, but both developed specialized models for more specific problems. The ANODEV approach we used is flexible and can be applied to a variety of gear comparison problems.

We analyzed the overall catch ratio with GLMs using both Poisson and negative binomial distributions (see full analysis on GitHub at https://github.com/tcwain/GearComparisonAnalysis2018/raw/master/MED_Gear_Analysis.pdf). We found the negative binomial model to be clearly preferable in terms of scale, stability, and approximate normality of residual deviances for all but two taxa (subadult Chinook Salmon and subadult Coho Salmon). In these two cases, the estimated negative binomial shape parameter θ (Table ) was high, suggesting a spatial distribution near CSR. For four species (Chum Salmon, Northern Anchovy, Pacific Herring, and Pacific sea nettle), the negative binomial model was far better than the Poisson model, but diagnostic plots showed that residual deviances still had notable departures from normality. This suggests that a different model of aggregation might be better—perhaps a zero‐inflated approach, such as that used by Minami et al. () and Burke et al. (). In fact, for the most strongly aggregated species (Pacific Herring), the iterative estimation algorithm for θ failed to converge, and we had to use a preliminary estimate of θ for that case.

In general, we found that the ANODEV approach provided a flexible framework for gear comparison studies that can be adapted to either paired‐sample or blocked designs with multiple gear types. Using a negative binomial distribution allowed a single model to estimate effects for a variety of species with different patterns of spatial aggregation ranging from nearly random (subadult salmon) to strongly schooling fishes (Pacific Herring and Northern Anchovy).

Methods used for the size‐selectivity analysis differed from those for estimating the overall catch ratios. For some of the species analyzed here, relatively few fish were measured, and those were often concentrated in a few hauls. Because of this, the data did not support the use of a complex model, and we instead used a standard size‐selectivity model fitted to the total size distribution for each gear across all hauls, combined with a double‐bootstrapping procedure to estimate error. Despite the differences in methods between the overall catch ratio analysis and the size selectivity analysis, results were consistent in showing the general effects of the MED. By using both methods, we were able to estimate overall calibration coefficients for our surveys and to also use the size‐specific patterns to improve our understanding of the biological factors affecting gear efficiency.

Operational Constraints

Long‐term research and monitoring surveys frequently face unanticipated operational changes due to equipment failure, change in available technology, vessel reassignment, or (as in our case) change in regulatory policy. Such changes result in the need for unplanned gear or vessel calibration studies, which are sometimes an afterthought and therefore often conducted with very limited resources.

Our initial conceptual design for this study was a random paired‐sample gear comparison with two vessels side by side. An initial statistical power analysis for that design suggested that for moderately aggregated species, such as juvenile salmon, detecting a 20% effect on catch would require about 100 samples (50 with each gear). We had insufficient budget or staff time to achieve that goal within a single season.

Beyond staff and budgetary constraints, we were also limited in sampling effort by the terms of our ESA permits for collecting salmon. Thus, from the start, sampling was a compromise between statistical rigor and operational reality. We were unable to charter a second vessel for side‐by‐side sampling, so we had to swap the MED in and out of the net between samples on one vessel. To do this efficiently, we settled on swapping the gear only every other tow. This resulted in a pattern of two samples with the MED followed by two samples without it. To maximize the number of samples in a day, we also minimized travel time by staying on a single station for a full day (or sometimes half a day). This changed the statistical design from a paired‐sample design to a blocked design.

Annual budgets and permitted catch limited the study to only a few days per year, so the study extended over 3 years. For these reasons, we had limited power to detect small to moderate effects; however, the observed effects in some cases were quite large (>50%). More extensive sampling would have provided narrower CIs in estimated catch ratios, which would have reduced uncertainty in correcting survey time series for use of the MED.

The cloud of limited annual sampling did have one silver lining: we were able to examine early results between cruises. Thus, we were able to recognize early that the original MED design resulted in substantial catch reduction for most of our targeted species of juvenile salmon. This, combined with the recent availability of inexpensive underwater video cameras, allowed us to experiment with various MED modifications during 2014 (Weitkamp ), resulting in the decision to invert the MED during 2015. Of course, because we did not directly compare the upward and downward orientations at the same time and location, we cannot ignore the possibility that the differences observed were due to environment effects on fish response rather than to differences between the two MED orientations.

Future Directions

Reducing the loss of target species may be possible through re‐design of the MED. During the 2014 cruise, we informally tested some simple modifications to the grate orientation, escape design, and flotation near the MED (Weitkamp ). Further research into fish behavior near the MED could lead to new ideas for improving its operation. In addition, tests of MED efficacy in excluding mammals (either with observations of live mammals or with inanimate surrogates) would allow evaluation of the tradeoffs between mammal protection and loss of target species. Engineering tests of MED performance are also warranted, especially flume testing (e.g., Gibson and Isaksson ) to determine the performance of the MED under different trawl speeds and different levels of cod‐end flow resistance. If the design is changed, further calibration studies will be needed to correct the CPUE time series before any modified MED is adopted for operational use. For future calibration studies, other sampling methods (e.g., recapture nets or cod end covers) should be considered instead of or in combination with paired‐gear methods.

Conclusions

Our study was developed in response to a conservation conflict: bycatch effects on marine mammals during salmon conservation research surveys. Although the overall impacts of research activities on marine mammals are small relative to other sources of mortality, the legal and policy implications are not. Our results show that there is no win–win solution to the conflict. Using the MED may protect some marine mammals but has a strong effect on retention of some salmon species and other small pelagic fishes.

In its original upward orientation, the MED reduces catch rates of most small pelagic fishes, and reductions were severe for Coho Salmon, Chum Salmon, Northern Anchovy, and Pacific Herring. With the downward orientation, the direction of effects was somewhat mixed, with moderate reductions in catch for our target juvenile salmon (Coho Salmon, Chum Salmon, and subyearling Chinook Salmon) but increases in catch for nontarget fish (subadult Coho Salmon, Northern Anchovy, and Pacific Herring).

The GLM ANODEV statistical approach proved useful and provides a potential method of correcting survey time series for the effect of the MED on target taxa, but the wide CIs on catch ratio estimates will substantially reduce the precision of adjusted CPUE estimates. In addition to the need for CPUE adjustments, using the MED will require substantial increases in effort needed to collect some target species for genetic, diet, and condition studies. Reducing these problems may be possible through re‐design of the MED.

Acknowledgments

This work could not have been accomplished without the help of many, including field science staff, crews of the FV Frosti and RV Ocean Starr, and land‐based support staff. We particularly thank Cheryl Morgan for her database wizardry, Ari Blatt for countless hours counting shadows in videos, and an anonymous reviewer for suggesting new methods for size‐selectivity analysis. Mark Lomeli, Mary Hunsicker, Kurt Fresh, JoAnne Butzerin, and two anonymous reviewers provided helpful reviews and comments on earlier drafts of the manuscript. We dedicate this work to our coauthor, Robert (Bob) Laurence Emmett (December 21, 1955–April 27, 2015). Bob was a trusted colleague, mentor, and friend who was taken from us much too early.

The scientific results and conclusions, as well as any views or opinions expressed herein, are those of the authors and do not necessarily reflect the views of the National Oceanic and Atmospheric Administration (NOAA) or the Department of Commerce. Reference throughout this document to trade names does not imply endorsement by the NMFS, NOAA. There is no conflict of interest declared in this article.