Comparing size selectivity and exploitation pattern of diamond-mesh codends for Southern velvet shrimp (Metapenaeopsis palmensis) in shrimp trawl fishery of the South China Sea

Sea trials for experimental fishing and population surveys

Sea trials were carried out onboard a commercial double-rigged shrimp trawler (name ‘Guibeiyu 96899’, engine power of 280 kW and total length of 38 m) in October 2019. Experimental fishing was conducted in the Beibu Gulf of the Northern SCS (Fig. 1), where is a traditional area for shrimp trawling and about 37 km from the coastline. To make sure a commercial fishing condition, towing duration and speed were kept mainly at 2 h and 3.5 knot. All experiments have been done during day and night continually, which is typical for the commercial fishery. In addition to the experimental fishing, three population surveys of the target shrimp were conducted onboard the same and similar trawler using the same trawl but with unselective codends in April 2017, November 2018, and March 2019, respectively.

Fishing gear and experimental set-up

The double-rigged trawls had a 38.7 m fishing circle (860 meshes in circumference with a mesh size of 45 mm) and total stretched length of 32.9 m. These trawls were equipped with a 28 m headline and a 36 m fishing line, to which an 80 m long chain (weighed 210 kg) was attached and acted as a ground-gear. Two sets of rectangle trawl doors (1.6 m² area and 250 kg weight) made of steel and wood were used to spread the trawls.

Fishing gear components were identical with commercial fishery except the codends, which were the main modifications focused in this study. Six diamond mesh codends, mesh sizes ranging from 25 to 54 mm, were designed using the dimension of the commercial codend, which had 220 × 25 mm meshes in the circumference and a stretched length of 4.8 m, as the baseline. The same net maker made all these codends. They were identical in both dimension (circumference and total stretched length) and material; the only differences were the mesh sizes. We termed these codend as D25, D30, D35, D40, D45 and D54, respectively, based on their mesh sizes, and detailed information about the specification and mesh openings were listed in Table 1. Following the recommendation of Wileman et al. (1996), we applied the covered codend method in the experiments. Compared with the experimental codends, the covers used were 1.5 times larger and longer and with a mesh opening about 12 mm (Table 1). To reduce the potential cover-effect, 12 flexible kites were attached around the covers (He, 2007; Grimaldo et al., 2009). Additionally, underwater video recordings (GoPro HERO 4 BLACK Edition) were applied to check whether the cover would mask the tested codend before and during the experimental fishing.

The double-rigged commercial trawler provided us an ideal workplace for the selectivity experiments, in which two separate codends could be tested simultaneously. In order to estimate the effect of mesh sizes on the selectivity of codends, we arranged the pairwised tests as: D25 vs. D30, D35 vs. D40 and D45 vs. D54, respectively. In other words, two pairwised codends were tested at a time for several hauls using the method mentioned above, then moved on to another pairwised test.

During the experimental fishing, all catch of the Southern velvet shrimp from the codend and cover were collected and sampled (if the catch number was huge), and frozen separately after the haul back process. Once we got back to the laboratory, all collected shrimp were counted and length measured to the nearest 1 mm for further selectivity analysis.

Size-selectivity estimation

The size selectivity of the tested codend for the target shrimp species was carried out separately with the method described below. Our experimental design enabled us to analyze the catch data as binominal data, as the target shrimp was either retained by the cover or the codend in a specific test. The retention probability of a specific codend for a given shrimp with length l in haul j can be expressed as r_j(l). The experimental value of r_j(l) can be easily calculated by the count number of the codend and the total number (codend + cover); while its theoretical value can be estimated using some parametric models. However, the value of r_j(l) could be expected to vary for the same codend between different hauls (Fryer, 1991). In order to account for this variation, we estimated the averaged values of r(l) by pooling data over all hauls for the tested codends, assuming that size selectivity performance of the tested codend in the experimental fishing can represent how the codend would perform in a commercial fishery (Millar, 1993; Sistiaga et al., 2010; Herrmann et al., 2016; Yang et al., 2021).

The averaged size selectivity mentioned above can be further described as r_av(l) (Herrmann et al., 2012). Four parametric models, Logit, Probit, Gompertz and Richards (Wileman et al., 1996), were used as candidate to test for r_av(l), in which v is a vector representing the parameters to be estimated. This analysis is conducted to select the best model to make experimental data most likely to be observed, with the assumption that the experimental data could be described by the selected model sufficiently. We minimized Eq. (1) to estimate the parameter v. This analysis was equivalent to maximize the likelihood of the experimental data in the form of number of shrimp, which is the length-dependent, caught by the codend (nR_jl) and those by the cover (nE_jl):

(1) $$-\sum _{\mathrm{j}=1}^{\mathrm{m}}\sum _{\mathrm{l}}\left\{\frac{n{R}_{jl}}{q{R}_j}\times \ln \left(r_{av}\left(l,v\right)\right)+\frac{n{E}_{jl}}{q{E}_j}\times \ln \left(1.0-r_{av}\left(l,v\right)\right)\right\}$$where the outer summation is over the m hauls carried out, while the inner summation is over length class l; qR_j and qE_j is the sub-sampling rate of the shrimp species length measured from the codend and cover, respectively.

A two-step procedure was applied: first, each of the candidate models was initially fitted in Eq. (1), then selected the best model with the lowest Akaike’s information criterion (AIC) values (Akaike, 1974); second, once the best model was identified for a given codend, a double bootstrapping technique was applied to calculate the Efron percentile 95% (Efron, 1982) confidence intervals (CIs) for the size selectivity and parameters, incorporating both within- and between-haul variation (Millar, 1993; Herrmann et al., 2012; Yang et al., 2021). Finally, the ability of the selected model to describe the experimental data sufficiently well can be evaluated according to the p-value, which represents the likelihood of obtaining at least as big a discrepancy between the modelled result and the observed data. Generally, the p-value from the selected model should not be less than 0.05; if it is not, p-value < 0.05, we should inspect the residuals to check whether the low p-value was due to structural problems of the model to represent the experimental data or if it was simply due to overdispersion in the data (Wileman et al., 1996; Yang et al., 2021).

Delta selectivity

In order to explore how the size selectivity of tested codends for the target shrimp change with increasing mesh sizes, we calculated delta selectivity, Δr(l), by the following expression:

(2) $$\mathrm{\Delta }r(l)=r_2(l)-r_1(l)$$where r₂ (l) is the size selectivity for codend 2 with a larger mesh size, and r₁(l) represents the size selectivity for codend 1 with a smaller mesh size. The Efron 95% CIs of Δr(l) was calculated according to the two bootstrap populations of results for both r₁(l) and r₂(l). Because they were obtained independently, we created a new bootstrap population for Δr(l) by the following expression:

(3) $$\mathrm{\Delta }r(l{)}_i=r_2(l{)}_i-r_1(l{)}_{i}i\in [1\dots 1000]$$where i is the bootstrap repetition index. As the bootstrap re-sampling of the two populations was random and independent, it is valid to generate the bootstrap population of results for the difference based on Eq. (3) (Herrmann, Krag & Krafft, 2018; Herrmann et al., 2019; Yang et al., 2021).

Estimation of exploitation pattern indicators

In addition to the estimation of size selectivity, it is essential to quantify how the tested codends with different mesh sizes perform under the same fishing population of the target shrimp, a specific scenario of population, nPop_l, was generated by pooling length data from both cover and coded over all hauls (Melli et al., 2019; Einarsson et al., 2021; Yang et al., 2021). In total, four different fishing population scenarios of the target shrimp species were generated. Applying the size selectivity predicted in the previous section, four exploitation pattern indicators, nP−, nP+, nRatio, and dnRatio (Eq. (4)), were calculated for each codend with a MCRS of the target shrimp. Specifically, as there is no formal MCRS promulgated for the Southern velvet shrimp in the studied area, we used its first matured total length of 7.0 cm (Liu & Zhong, 1988) as a MCRS to calculate the indicators.

(4) $$\begin{array}{c}\begin{array}{c}\begin{array}{c}nP-=100\times \frac{\sum _{l<MCRS}\{r_{codend}\left(l\right)\times nPo{p}_l\}}{\sum _{l<MCRS}\{nPo{p}_l\}}\\ nP+=100\times \frac{\sum _{l\ge MCRS}\{r_{codend}\left(l\right)\times nPo{p}_l\}}{\sum _{l\ge MCRS}\{nPo{p}_l\}}\\ nRatio=\frac{\sum _{l<MCRS}\{r_{codend}\left(l\right)\times nPo{p}_l\}}{\sum _{l\ge MCRS}\{r_{codend}\left(l\right)\times nPo{p}_l\}}\end{array}\\ dnRatio=100\times \frac{\sum _{l<MCRS}\{r_{codend}\left(l\right)\times nPo{p}_l\}}{\sum _l\{r_{codend}\left(l\right)\times nPo{p}_l\}}\end{array}\end{array}$$where r_codend(l) is the size selectivity from the tested codend, and nPop_l is the size structure of fishing shrimp population with length class l. The indicator nP− and nP+ is the percentage of shrimp with length below and above the MCRS retained by the tested codend, respectively. For a tested codend with good selective properties, nP− value is expected to be close to 0 while nP+ value close to 100%. The indicator nRatio represents the landing ratio of shrimp with length below and above the MCRS. The indicator dnRatio is the total percentage of shrimp with length below the MCRS retained by the tested codend. To have good selective properties, nRatio and dnRatio for the tested codends is expected to be the lower the better. Again, we used the double bootstrapping approach mentioned above to estimate the Efron percentile 95% CIs for the indicator values.

The data analysis mentioned above, including size-selectivity estimation, delta selectivity and estimation of exploitation pattern indicators were conducted using the selectivity software SELNET (Herrmann et al., 2012).

Experimental and survey data

In total, 47 valid hauls were carried out; eight hauls for each codend, except the D45 codend for which seven hauls were conducted (Table 2). The averaged duration was about 130 min, with a range of 118 to 156 min, and the water depth was mainly 16 m, ranging from 12 to 24 m, in the fishing grounds. Among the species caught, Southern velvet shrimp was present in all hauls and was the most dominant species in terms of quantity. We obtained sufficient number of the target shrimp to be included in selective analysis, 3,334 in total, 1,815 individuals from the tested codends and 1,519 from the covers. Sub-sampled ratios for the target species ranged from 0.2 to 1.0, depending on the catch amount of the specific haul.

The fishing population scenarios were generated by pooling length distribution of all Southern velvet shrimp in four different sea trials (Fig. 2). These population scenarios were based on sufficient number of length measurement for the target shrimp. In addition to the number of individuals in the experimental fishing mentioned above, there were 2,632, 1,726 and 1,860 individuals of length measurement in the population survey of 2017, 2018 and 2019, respectively.

Size selectivity

Based on the AIC values of four candidate models, the best model was selected as: the Gompertz for the D25 codend, the Richards for the D30 codend, the Probit for the D35 and D40 codend, and the Logit for the D45 and D54 codend, respectively (Table 3).

The fit statistical results demonstrated that all selected models were able to express the experimental data well, except the D25 codend which a p-value < 0.05 was obtained. For this codend, however, as the selectivity curve represented the main trend of the experimental data well (Fig. 3A), we concluded that this low p-value was due to overdispersion in the fishing data.

As the mesh sizes increased, the selective parameters of the tested codends, both L50 and SR, for the target species represented an increasing trend. For instance, L50 and SR of the D25 codend was 5.47 and 0.89 cm, respectively, whereas the relative value increased to 7.72 and 4.01 cm for the D54 codend, respectively (Table 4). The codends with the largest mesh sizes, D45 and D54, had significantly larger L50 than the codends with smaller mesh sizes, the D25, D30 and D35 codend, respectively, while the SR of the D45 and D54 codend was significantly larger than those of the D25 and D30 codend, respectively. Additionally, the CIs in selective parameters became wider for the tested codends as the mesh sizes increased, especially for the D54 codend. The results from selectivity curves showed a similar fishing pattern. As the mesh sizes of tested codends increased, the retention probability for shrimp with a MCRS length decreased. For example, the retention probability of shrimp with a MCRS length in the D25 codend was larger than 95%, while this value dropped to 52% for the D40 codend and below 41% for the D45 and D54 codend, respectively.

Delta selectivity

The results of delta selectivity demonstrated that applying codends with larger mesh sizes would decrease the retention probability for the studied species, and most of these differences were statistically significant and length-dependent. For instance, comparing with the D25 and D30 codend, the D35 and D40 codend had significant lower retention probability for shrimp with an approximate length range of 5.5 to 10.0 cm (Fig. 4). The D45 codend significantly retained fewer shrimp with lengths above 5.4 cm, 5.8 cm and 6.3 cm than the D25, D30 and D35 codend, respectively (Fig. 5). A similar tendency was obtained for the D54 codend by comparison with the codends with smaller mesh sizes. The effect of four comparisons was not significant including D30 vs. D25, D40 vs. D35, D45 vs. D40 and D54 vs. D45, respectively.

Exploitation pattern indicators

The exploitation pattern indicators showed that catch efficiency, both undersized (nP−) and marketable size (nP+), of codends used for the target shrimp decreased with the mesh sizes increased. In all fishing population scenarios, for instance, the D25 codend retained more than 43% of undersized shrimp (nP−), the ratios were above 29% for the D30 and D35 codend, respectively, whereas the percentages dropped to less than 29% when the mesh sizes used larger than 35 mm (Table 5). Applying codends with larger mesh sizes, however, would compromise the catch efficiency of shrimp with marketable sizes. For example, the D25 and D30 codend retained more than 95% of legal size shrimp (nP+), and the D35 codend retained above 75%. In comparison, about 60% of shrimp with marketable size was caught by the D40 codend, and the ratios were nearly 50% for the D45 and D54 codend. Some of these differences mentioned above were statistically significant. The results of nRatio and dnRatio varied in different fishing population scenarios for the same codend, and differences between codends were not statistically significant in the same scenarios. For instance, the discard percentage of shrimp caught (dnRatio) was below 30% for population scenario in 2017, the relative values were all above 76% for all codends used in population scenario of 2018.