Sampling for estimation of catch composition in Bering Sea trawl fisheries

Executive Summary

Management of groundfish fisheries in Alaska is based on annual, seasonal, or fishery- and vessel-specific catch limits. Limits include both total allowable catch quotas for major species and incidental catch limits for many non-target species, including prohibited species such as Pacific halibut (Hippoglossus stenolepis).  Fisheries are managed in near-real time based on industry reports and data collected by at-sea observers.  Estimates of both the total catch for sampled hauls and the species composition of individual hauls are based on randomly selected samples collected by observers.  The precision and accuracy of observer sampling are, therefore, of considerable importance to both industry and regulators of these fisheries. Accuracy of sample composition estimates is particularly critical where total catch is estimated for individual vessels or small fleet sectors; variability in estimates can have large effects on catch accounting.

We present results of two studies conducted in the eastern Bering Sea aboard commercial trawl catcher/processors.  These two studies had three common goals:

1) To evaluate alternatives for selection of catch composition samples,
2) To check for possible biases associated with sample selection, and
3) To estimate the precision of catch composition estimates based on selected samples.

The first study, conducted in 1999 aboard the FV American No. 1, used modified standard observer sample collection methods and looked for evidence of mechanical sorting or stratification of species during net retrieval and catch handling aboard the vessel. Samples for this study were selected systematically throughout the haul. These sample-based catch estimates were compared to catch estimates based on processed catch product for targeted species, and to censuses of catch for non-target species.  For this study, target species included walleye pollock (Theragra chalcogramma), Pacific cod (Gadus macrocephalus), yellowfin sole (Limanda aspera), flathead sole (Hippoglossoides elassodon), and Alaska plaice (Pleuronectes quadrituberculatus).  Non-target species included in the study were Pacific halibut, skates (Raja sp. and Bathyraja sp.), Tanner crab (Chinoecetes bairdi), snow crab (Chinoecetes opilio), and red king crab (Paralithodes camtschaticus). 

The second study, aboard the FV Seafisher in 2005, tested an automated catch sampling and monitoring system as a means to limit mechanical sorting and to remove potential bias from the sample selection process. The automated sample selection system used a factory-based computer to determine when the sample should be selected, and then diverted catch from the processing line to the observer sample station. Samples were collected from the haul using a simple random sampling design.  Catch estimates based on sampling results (sample-based catch estimates) were compared to censuses of catch for selected non-target species and to the difference between the total haul weight (flow scale) and censused non-target catch weight for the target species (yellowfin sole).  Non-target species included in this study were Pacific halibut, arrowtooth flounder (Atheresthes stomias), Kamchatka flounder (Atheresthes evermanni), Pacific halibut, and eelpout species (Family Zoarcidae, all species).  Both studies provided information on the variability of catch composition estimates between hauls and within multiple samples of each haul.

A simulation study was conducted based on data collected during the Seafisher research to examine the effects of sampling fraction on estimates of species composition. A simulated haul (28 metric tons) was constructed consisting of six species of fish, essentially mimicking the five major species encountered in the Seafisher data set, and a last species that represented all other species. Fish were randomly assigned to a sample until the sample achieved the target weight of fish. Since only whole fish are included in the sample, the weight of fish in the sample varied. Estimates of catch based on the simulated samples were compared with the true catch (simulated haul total for the species). Bias and variance of the estimates was evaluated.

Results

• Measurements of codend volume provided a reasonable approximation of total catch weight in both studies. When the volume of the codend was measured, the volumetric estimate was generally within 15% of the weight measured by the vessel’s flow scale.
• The automatic sample collection and electronic monitoring (EM) systems tested on the FV Seafisher performed well, but two concerns were identified.  First, since the total catch size was unknown prior to sampling, selection of a fixed number of random samples was difficult.  When the initial volumetric estimate of total catch was an overestimate, sometimes a smaller number of samples were collected than was desired. When total catch was initially underestimated, the final portion of the catch was not included in the random selection.  Secondly, random selection of samples sometimes led to samples that were too close together to be efficiently processed by the observers. Both of these concerns could be addressed by systematic sampling with a random start point.  Actual weight of samples diverted by the system varied somewhat from the target 100 kg, primarily due to accumulation of fish at the inclined conveyor belt.  The use of EM, in particular, appears to have the potential to increase compliance with catch-sorting protocols and smooth the sampling process.
• While both studies fished with bottom trawl gear on the central Bering Sea shelf, fishing methods and the overall composition of catch differed between the two studies.  Catch in the American No. 1 study was a mixture of yellowfin and flathead sole, walleye pollock, and Pacific cod. Catch in the Seafisher study was dominated by the targeted species, yellowfin sole. These differences produced differences in the variability of the catch composition between hauls in each study.  The coefficients of variation (CVs) of product or census-based catch proportions for the Seafisher study were 7% for yellowfin sole and 49-63% for the three rare species. For the American No. 1 experiment, the four dominant species had CVs of 55-95%, while CVs for crabs and Alaska plaice were on the order of 100-200%.
• Estimates of species composition in both studies were calculated for sample sizes of 100, 300, and 600 kg. While the overall means of species proportions estimated from the samples tended to be very close to the product or census-based means, the range and variability of the sample estimates differed substantially between studies and between species. Overall CVs of sample-based estimates of species proportion for the American No. 1 study were high; in the range of 60-80% for the dominant species and over 100% for the rare groups. Species proportion estimates from the Seafisher study showed similar patterns. The single dominant species in the Seafisher study was well represented even at 100 kg sample sizes; the overall CVs for this species were 8-11%. The three rare species groups in this study, however, showed wide ranges in estimates of sample proportion and had overall CVs of over 100% at even the largest sampling fraction, with CVs for 100 kg samples of 300-600%. For all species, increasing the size of the sample had little effect on the overall mean of the sample estimates but markedly reduced the range of individual estimates of the species composition. Overall CVs for each species group decreased with increasing sampling fraction.  This effect was slight for the predominant species, but pronounced for the less common species, especially Pacific halibut. 
• For each haul, the difference between census-based estimates and sample-based estimates was calculated; the frequency distribution of these differences was examined to check for any sampling bias. In general, these distributions fall into three distinct groups depending on the overall abundance of the species being sampled.  For dominant species in both studies (e.g., yellowfin sole, walleye pollock), the distribution of the differences was symmetric around zero for all sample fractions.  Rare species groups in both studies (Pacific halibut, skates, and Kamchatka flounder) show a distinctive, strongly asymmetrical pattern in the differences between sample and census-based estimates. The strong skewness in the distribution caused the majority of samples to underestimate the proportion of these species as zero (when none of the rare species appear in the sample), but a few samples to overestimate the proportion by a large extent (when one or few of the rare species is present in the sample). This skewed distribution can be expected to occur when the overall average proportion for a species or species group is very small. 
• Results of the simulation study were consistent with the two field studies. Means of the sample-estimated haul weight for each species were close to the true values and did not change substantially with sampling fraction. The precision of the sample estimates did, however, change substantially with sampling fraction.  The variability of the sample estimates decreased with increasing sample fraction for all of the studied species; the rate of decrease was fastest at the lowest sampling fractions and for rare species.  Frequency distributions for rare species were strongly right-skewed at low sampling fractions but became progressively less skewed at increasing sampling fractions. Distributions for rare species did not become symmetric, however, until the sampling fraction exceeded 37% of the total haul weight.  
• Comparison of variability between product and census-based estimates and within-haul sample estimates indicate that between-haul variability is the dominant variance component for target species in these studies. For rare species, however, variability of sample estimates was much higher than variance of product and census-based estimates, indicating substantial within-haul variability due to sample selection.
  

Conclusions

Results of these field studies indicate that existing observer sampling protocols based on 300 kg standard samples provide good estimation of catch composition for target and common groundfish species.  While there was evidence of slight stratification of species composition within the trawl net, sample estimates of proportion were generally in good agreement with production and census-based species proportion estimates. Even small samples (100 kg) provided estimates of catch composition for predominant species that were close to production and census-based species proportion estimates. 

The most important observation from both studies was the pattern revealed in estimation for rare species, including Pacific halibut.  Where management of a fishery includes catch limits on prohibited or non-target species, the poor precision of the estimates for rare species has potentially serious consequences. If precise estimation of catch of rare species is desired, large sampling fractions are needed to provide haul-specific estimates with small variance. Where large sampling fractions cannot be achieved, then combined estimates over a number of hauls are needed to obtain precise estimates of catch for the combined hauls.  The relative importance of competing management goals and the eventual use of observer data in management will need to be explicitly considered in structuring of observer data collection programs.

View Online  (.pdf, 2.33 MB).