Murray: Magnitude and distribution of sea turtle bycatch in the sea scallop dredge fishery 



673 



Data sources 



Observer data Observers were placed on randomly 

 selected vessels fishing in the controlled areas to record 

 the bycatch of turtles and other protected species. From 

 May to December in 2001 and 2002, observers sampled 

 11% of the commercial fishing effort in the Hudson 

 Canyon region, and in October 2001, 16% of the effort in 

 Virginia Beach. No trips were observed in the Virginia 

 Beach region during 2002 because of low commercial 

 fishing effort in the area. Observers were on- and off- 

 watch on an irregular schedule throughout a 24-hour 

 period, observing on average 65% of the hauls on a trip. 

 When a dredge was hauled on board, observers recorded 

 the haul location, time, depth, tow speed, tow duration, 

 number of dredges observed, and the presence or absence 

 of turtle bycatch. In 2001, observers identified 20% of the 

 turtles that came aboard as loggerhead sea turtles but 

 were unable to identify the remaining 80%. As a result 

 of improved observer training (NMFS 200.3), observers 

 identified 88% of the turtles as loggerhead sea turtles 

 in 2002, but they were unable to identify the remain- 

 ing 12%. Given that observers document the loggerhead 

 species most commonly in the Mid-Atlantic area, and 

 that all sea turtles positively identified were loggerhead 

 sea turtles, bycatch estimates in this analysis are con- 

 sidered to be those of loggerhead sea turtles. Although 

 some turtles may have been released alive or injured, 

 this analysis does not differentiate between live, dead, 

 and injured animals. 



Fishing effort data Under the 1982 Atlantic Sea Scallop 

 Fishery Management Plan, all vessels targeting scallops 

 must complete a vessel trip report ( VTR) log (as of 1994) 

 indicating area fished, kept and discarded catch, and 

 fishing effort. These data were used to estimate the total 

 fishing effort of the fleet. In calculating fishing effort, 

 one unit of effort equals a single dredge haul because 

 vessels may fish one or two dredges simultaneously on 

 each haul. Because a preliminary analysis showed that 

 tow duration or dredge length does not significantly 

 affect the probability of turtle capture, dredge haul effort 

 was not standardized for these two variables. All VTR 

 trips from May to December in the controlled areas were 

 used in the analysis. Because completion of vessel trip 

 reports is mandatory and trips to the controlled areas 

 were closely monitored, it was assumed that the VTR 

 data represented 100% of total fishing effort. 



Sea surface temperature Sea surface temperature at 

 each position reported in the observer and VTR data- 

 bases was extracted from NOAA AVHRR (advanced 

 very high resolution radiometer) coastwatch satellite 

 images. A Visual Basic (Microsoft Corp., Redmond, WA) 

 routine was used to extract temperatures from 7-day 

 composite images (3 days forward and backward from 

 the haul date), by using a 3x3 cell window at 1-km 

 resolution. Therefore, a 9-km 2 area of coverage around 

 each coordinate position was used to extract sea surface 

 temperature. Within the 3x3 cell search radius, the pixel 



representing the warmest temperature was used to avoid 

 temperatures affected by cloud coverage. 



Data analysis 



Missing temperature data Sea surface temperature 

 values could not be obtained for 33% of the VTR data 

 and 10% of the observer data because of either missing 

 coordinate positions on the VTR logs or bad satellite 

 images. For these fishing events, sea surface tempera- 

 ture was predicted by using a linear regression based 

 on year, month, and area. For the observer data, area 

 was defined as either Hudson Canyon or Virginia Beach 

 access areas (r 2 =0.88). For the VTR data, the vessel's 

 home state served as a proxy for area fished because 

 most of the missing temperature values were due to 

 missing coordinate positions (r 2 =0.86). 



Modeling approach Generalized linear model (GLM) 

 and generalized additive model (GAM) fitting techniques 

 were used to understand and predict bycatch rates of sea 

 turtles in relation to environmental variables, fishing 

 practices, and gear characteristics in the commercial 

 sea scallop fishery. Unlike classic linear regression 

 models, GLMs and GAMs allow for nonlinearity and 

 nonconstant variance structures in the data (Guisan 

 et al., 2002). GAMs differ from GLMs in that smooth 

 functions replace the linear predictors in GLMs (Hastie 

 and Tibshirani, 1990). Smooth functions, or "smoothers," 

 summarize the trend of a response measurement as a 

 function of multiple predictors (Hastie and Tibshirani, 

 1990) and therefore some form of parametric relation- 

 ship between the response and explanatory variables 

 is not assumed (Guisan et al. 2002). Both frameworks 

 have been used to model abundance or probability events 

 as a function of environmental variables (Frost et al., 

 1999; Denis et al., 2002; Guisan et al., 2002; Hamazaki, 

 2002). 



A modeling approach to estimate bycatch of sea tur- 

 tles in the sea scallop dredge fishery was preferred over 

 the ratio method (Cochran, 1977) that has been used 

 to estimate bycatch of marine mammals and turtles 

 in other fisheries (Epperly et al., 1995; Rossman and 

 Merrick, 1999). With the ratio method, the observed 

 number of sea turtles divided by the observed effort is 

 used to calculate a bycatch rate, and this rate is then 

 multiplied by total commercial fishing effort to derive 

 a bycatch estimate. Bycatch data in the sea scallop 

 dredge fishery violate the underlying assumptions of the 

 ratio method (Cochran, 1977), largely because sea turtle 

 bycatch is binomially distributed with a nonconstant 

 variance. An analyis of binary response data derived 

 from a statistical model allows bycatch rates to be pre- 

 dicted by using factors that account for variability in 

 bycatch. Moreover, stratifying bycatch rates according 

 to these factors will reduce variability in total bycatch 

 estimates. For the sea turtle data analyzed in the pres- 

 ent study, the GLM approach provided a more accurate 

 and less biased mortality estimate than that derived 

 using the ratio method. 



