A Machine Learning Technique for Deriving the Optimal Mesh Size of a Gizzard Shad (Konosirus punctatus) Gillnet

Abstract

Gizzard shads are facing a continual decline in population, necessitating the implementation of selective gear design for effective resource management. This study aims to prevent the bycatch of young gizzard shads, a non-target fish species, and to derive mesh sizes appropriate for fishery management. Experimental fishing (n = 11) was conducted by manufacturing gillnet fishing gear with different mesh sizes (50.5, 55.1, 60.6, and 67.3 mm) in the coastal waters of the southern Gyeongsang Province. Two methods were employed to estimate the appropriate mesh size of the shad gillnet as follows: firstly, by analyzing the selectivity curve based on body length data; secondly, by developing a complex machine learning model considering biological and economic factors. Model 1 was constructed using mesh variables to classify the score groups. As a result of this study, the total length with a 0.5 gillnet selection ratio. which was estimated to be 179.3, 195.6, 215.1, and 238.9 mm for the 50.5, 55.1, 60.6, and 67.3 mm mesh sizes, respectively. In Model 1, a mesh size of 57.85 mm or less was determined as the most appropriate mesh size. Therefore, considering both biological and economic aspects, shad gillnets should have a mesh size in the 50.5 to 55.1 mm range.

1. Introduction

Most fishing gear is designed to target one or a few species. Therefore, a specific target fish is selected, the gear’s material and structural characteristics are modified, and various fishing methods are employed. It is widely known that the size selectivity of commercial fishing gear is crucial in fishery management to maximize yield and protect juvenile fish [1,2]. Moreover, fishing gear may be used as research tools for monitoring the length and distribution of the stock using the gear’s size selectivity to adjust the length distribution of catches [3].

Gillnets work in the following three ways: sticking, entangling and confining the gills of the target species into the net. Gillnet gear is widely used to identify the species and length characteristics of catches [4,5,6,7,8,9,10,11,12,13]. Furthermore, gillnet materials and fishing methods differ depending on the country, region, and target species. In the Republic of Korea, coastal gillnet fishing was practiced by 18.7% (12,262) of coastal fishing vessels (65,531) in 2022 [14]. There are four main gillnets as follows: set gillnets for fish living in the lower layer [6,15,16], drift gillnets for fish living in the upper layer [8], encircling types of gillnets for schooling fishes [5], and trammel gillnets for entanglement [9].

The selectivity of gillnets is currently studied through the modification of the gear material [4,9] or previous studies focused on different target species mesh sizes [2,5,7,8,10,11,12,13,15,16,17].

In Ghana, selectivity curves for five commercially valuable species (Sarotherodon galilaeus, Oreochromis niloticus, Coptodon zillii, Clarias gariepinus, and Auchenoglanis occidentalis) were estimated at different mesh sizes of gillnets in the Tono Reservoir [16]. Additionally, in Queensland, Australia, the research investigated the selectivity of gillnets using different mesh sizes for two elasmobranch species (shark and ray) [17].

Abroad, studies have mainly focused on investigating gear selectivity for multiple species using the same gear. However, in domestic research, studies have primarily examined gear selectivity for a single species in specific regions. Park et al. [18] investigated selectivity using gillnets with different mesh sizes for Spinyhead sculpin (Dasycottus setiger) in the eastern sea of the Republic of Korea. An et al. [8] conducted research on gear selectivity for Pacific herring (Clupea pallasii) using gillnets along the coast of the East Sea.

Gizzard shad (Konosirus punctatus) is one of the most commercially essential fish species in the Republic of Korea [19]. However, the gizzard shad fishery displays high between-years variability, with sharp declines in catches from 11,002 MT in 2012 to 6649 MT in 2022 [14]. According to fisheries production statistics from 2012 to 2022, excluding items classified under miscellaneous fisheries, gizzard shad were caught in a total of 26 fisheries. It was found that 31% of the catches were from coastal purse seine and over 21% from coastal gillnet [14]. These statistics show that the gizzard shad is primarily caught using coastal purse seines and coastal gillnets. However, when it comes to purse seine gear, the focus is not on enhancing gear selectivity by altering mesh sizes or materials. Rather, the emphasis lies on developing optimal gear, such as modifying mesh sizes or gear materials to minimize hydrodynamic resistance during fishing operations and improve the sinking force of the gear [20]. Therefore, to enhance the selectivity and protect gizzard shad resources, it is necessary to conduct investigations into multiple sectors involved in catching the gizzard shad. However, considering time and cost constraints, it is crucial to prioritize research on gear selectivity in coastal gillnet fisheries, where the catch ratio of the gizzard shad is high, and formulate fishery management plans accordingly.

Studies on the gizzard shad include research on their reproductive biology [21], age composition and breeding season [22], and genetic diversity and population structure [23]. However, few studies have considered selectivity targeting the gizzard shad despite the importance of size selectivity data in managing fishery resources. Thus far, gizzard shad selectivity using encircling gillnets has been studied in the eastern coastal region of the Republic of Korea [5]. Notably, approximately 60% (2012–2022, [14]) of the gizzard shad landing is caught in the southern coastal region of the Republic of Korea. Therefore, it is appropriate to increase the gizzard shad fishery selectivity in this region.

In deriving an appropriate mesh for resource management, determining mesh selectivity is considered the most appropriate method. This is based on the maturity of fish, with anything above a 50% selection range considered suitable. However, Polet and Depestele [24] argue that ensuring sustainable fishing requires the consideration of both environmental effects and the cost-efficiency of fishing techniques and gear development. Therefore, it is necessary to derive the appropriate mesh size by considering a combination of biological aspects (selectivity for juveniles and non-target species) and economic factors (fishery efficiency). A model developed to make decisions considering these complex factors was developed in the fisheries sector [25]. In this model, a decision tree was developed to assess whether fishers should engage in fishing operations, taking into account various conditions such as weather conditions in fishing grounds, fish prices, and the potential for high catches in specific fishing areas [25]. In particular, the model utilized here, decision tree analysis, classifies or predicts by expressing decision rules in a tree structure, and it is comparatively easier to interpret than other machine learning models.

This study aims to prevent the bycatch of young gizzard shads and to determine the appropriate mesh size for the sustainable exploitation of fisheries. Test fishing operations were conducted by manufacturing gillnet fishing gear with different mesh sizes (50.5, 55.1, 60.6, and 67.3 mm) in the coastal waters of the southern Gyeongsang Province. The species, total length, and weight composition of catches caught in these gillnets were investigated during this study. The mesh selectivity curve was estimated using the SELECT model based on the investigated results. Furthermore, an analysis was performed to derive an appropriate mesh by applying a decision tree model and machine learning techniques.

2. Materials and Methods

2.1. Test Gear Design and Fabrication

Interview surveys targeting fishermen who engage in gizzard shad gillnet fishing were conducted to determine the usual mesh size, which is usually between 50 and 53.2 mm. Four types of gizzard shad gillnet gear with different mesh sizes (50.5, 55.1, 60.6, and 67.3 mm) were tested (

Table 1. Fabricated Konosirus punctatus gillnet test gear specifications.

Mesh Size (mm)	Height (Mesh Number)	Length (Mesh Number)	Number of Floats	Length between Floats (cm)	Number of Sinkers	Length between Sinkers (cm)	Mesh Number between Floats	Mesh Number between Sinkers	Float Line Length (m)	Sinker Line Length (m)
50.5	100	1182	35	87	101	30	35	12	29.6	30
55.1	92	1083	35	87	101	30	32	11	29.6	30
60.6	83	985	35	87	101	30	29	10	29.6	30
67.3	75	887	35	87	101	30	26	9	29.6	30

). The mesh cloth consisted of a nylon monofilament of 1.5 mm in diameter. The complete panel was 29.6 m long and 4.4 m in height (width) (Figure 1). Based on the hanging ratios of conventional gizzard shad gillnets, the floater line hanging ratio was 49.5%. Twenty openings were measured for each test gear type to obtain the test gillnets’ mesh opening sizes, and the average values were 50.5, 55.3, 60.6, and 67.7 mm, respectively.

2.2. Test Fishing Operation and Survey

The test fishing operations were conducted over four rounds from September to October 2016 and August to November 2017 off the coast of Goseong-gun, Gyeongsangnam-do Province, Goseong Bay (Figure 2), from a coastal gillnet fishing vessel (2016: 2.15 tons, 190 PS; 2017: 2.57 tons, 200 PS). During test fishing, the four gillnet sizes were arranged sequentially (Figure 3), twice, and by mesh size [26]. In addition, eight panels were configured into each set to test for potential differences in catch efficiency. For the test fishing operations, the four gillnet sets were shot 1 to 2 h before sunrise and hauled after sunrise. Shooting took about 30 to 60 min, and hauling took about 120 to 150 min. Shooting was performed by manpower while the fishing boat was moving forward at a slow speed. And hauling was conducted by a net hauler while moving the boat in the direction of the net. A local gillnet fishing vessel was chartered for test fishing. Three test gear sets (24 panels = 3 sets × 4 mesh sizes × 2 panels) were used during the first round in 2016; four sets (32 panels = 4 sets × 4 mesh sizes × 2 panels) were used in the second to fourth rounds during 2016; and four sets (32 panels = 4 sets × 4 mesh sizes × 2 panels) were used in 2017. Shooting and hauling were performed thrice per test operation round in 2017, excluding the seventh round (two times).

Throughout the test period, 160 panels were used for each test gear type (

Table 2. Fishing operations and gear conditions.

Year	Test Rounds	Fishing Operation Period	Shooting/Hauling Events	Test Gear Panels	Panels per Test Gear	Water Depth (m)	Water Temperature (°C)
2016	1	9.7–9.9	3	72	18	4.3–7.5	25.0–26.5
	2	9.22–9.24	3	96	24	5.2–8.4	23.4–23.9
	3	10.12–10.14	3	96	24	3.5–9.5	21.2–22.0
	4	10.26–10.28	3	96	24	3.5–11.0	18.3–21.3
	Subtotal	12 days	12	360	90	3.5–11.0	18.3–26.5
2017	1	8.9–8.11	3	96	24	5.0–10.0	29.0–29.5
	2	8.23–8.25	3	96	24	8.0–14.0	28.5–29.4
	3	9.6–9.8	3	96	24	4.5–11.0	24.4–24.8
	4	9.19–9.21	3	96	24	5.0–6.5	23.7–24.5
	5	10.11–10.13	3	96	24	7.0–12.0	22.0–26.0
	6	10.25–10.27	3	96	24	5.0–7.0	18.0–18.6
	7	11.14–11.15	2	64	16	5.5–10.5	15.3–15.7
	Subtotal	20 days	20	640	160	4.5–14.0	15.3–29.5
Total		32 days	32	1000	250	3.5–14.0	15.3–29.5

). The depth of the fishing grounds was 3.5–14.0 m, the water temperature was 18.3–29.5 °C, and the test gillnets immersion time was approximately 1–3 h (Table 2). Catches were classified by the tested gillnet mesh size during hauling, gathered in collection nets and returned to the port for measurement. Caught individuals were measured with a measuring board to record the fish’s total length, fork length, body weight, and sex. For the gizzard shad, this studies main target species, a fish measuring board was used to measure the total length, fork length, and body height in millimeters, and the body weight was measured in grams using an electronic scale (CAS SW-02, Korea).

2.3. Mesh Selectivity Curve Estimation

The Ishida [27], Holt [28], Sparre and Venema [29] and SELECT methods [2,30,31,32] were used to estimate the gillnet gear’s mesh selectivity. The Kitahara method, which displays the mesh selectivity curve as one master curve, was employed to estimate the mesh selectivity curve of gillnets [33,34,35] for further analysis. Fujimori and Tokai’s [33] polynomial method provided the mesh selectivity curve as a single reference curve with a left-right symmetrical shape for a quadratic function and a left-right asymmetrical shape for a cubic function of each curve. To determine which type was more appropriate for the selectivity curve, the unbiased estimator value for the error analysis was calculated for both quadratic and cubic functions. The cubic function that produced a smaller unbiased estimator value was then utilized in the analysis [35].

2.4. Statistical and Machine Learning Analysis

Statistical analysis was conducted using Python’s statistical package, utilizing the variables (month, mesh, mature, bycatch, CPUE, and score) described in

Table 3. Statistical analysis and decision tree variables.

Variable	Description
Month	Test operation month
Mesh	Mesh size
Mature	Maturity was judged based on the gizzard shad’s total and crotch lengths. The minimum mature length was 15.8 cm, set to 1 if it was less than the mature length and 0 if it was longer than the mature length.
Bycatch	The bycatch rate of fish from each net was calculated (other catches excluding shad/total catch × 100).
CPUE (catch per unit effort)	The standard was set as the number of fish per tested fishing gear, and the catch per unit effort in each test fishing operation was calculated (total catch in one fishing operation/number of fishing gear per fishing gear).
Score	A score was made using gizzard maturity, bycatch rate, and CPUE data to determine the optimal net for gizzard shad fishing. Excluding mature, bycatch and CPUE were recorded to 0 and 1 based on the scoring average. However, there is no standard value for bycatch rate or CPUE for optimal fishing with shad gillnets, so the average was derived from the 2016–2017 fishing season as the standard. Thus, the maturity of gizzard shads was based on mature data. Bycatch calculated the average (22.29%) of the bycatch rate in all test operations and scored it as 0 if it was below the average and 1 if it exceeded the average. CPUE is calculated as the CPUE average (8.27) in all test operations and is scored as 0 if it is above the average and 1 if it is below average. The score ranged from 0 to 3, with the standards set as 0 = very good, 1 = good, 2 = bad, and 3 = very bad.

to perform correlation analysis between each variable (p < 0.05). In addition, to develop a composite model that considered both biological and economic aspects in deriving the optimal mesh size for the gizzard shad gillnets, a decision tree model, which represented decision rules in a tree structure, was chosen from among machine learning models for its relatively easy interpretability. The decision model was structured similarly to Figure 4, starting from the root node and continuously branching based on whether the test requirement was Yes (True) or No (False), ultimately leading to the final nodes [36]. Therefore, the optimal classification criteria could be derived by considering the test requirements from the root node to the final load (leaf node). These splits occurred in a way that decreased impurity, and in this study, the Gini impurity was used to evaluate information gain.

For machine learning analysis, data on the season (month), body length (total and fork length), and catch per unit effort (CPUE) were used with data from test fishing preprocessed. Among the 4448 data obtained by test fishing, gizzard shad catch information and data with missing values were removed from the analysis, and statistical and machine learning analysis was performed on 2814 data using Python 3.11.

Table 3 presents the variables used for statistical analysis and machine learning. There were two machine learning models to classify score groups as follows: Model 1 was constructed using mesh variables, and Model 2 was constructed using mature, bycatch, and CPUE. In particular, the construction of Models 1 and 2 involved various combinations of variables to enhance the predictive performance of the models and ultimately determine the appropriate variables. Here, in Model 1, the score variable, which was scored to ultimately determine the effectiveness of the gear, was set as the target variable (dependent variable), with mesh size intervals set as independent variables. Model 2 was constructed to identify the variable elements that most influenced the score variable.

3. Results

3.1. Bycatch Rate

When comparing the total catch regardless of fish species in shad gillnets with different mesh sizes (

Table 4. Catch numbers and weight by species and test gear mesh size.

Mesh Size Species	50.5 mm		55.1 mm		60.6 mm		67.3 mm		Total
	No. of Species	Weight (g)	No. of Species	Weight (g)	No. of Species	Weight (g)	No. of Species	Weight (g)	No. of Species	Weight (g)
Trichiurus lepturus	1	95	1	77			1	142	3	314
Acanthopagrus schlegeli	3	175			1	73			4	248
Muraenesox cinereus	2	754	2	455	2	998			6	2207
Paralichthys olivaceus			3	122	1	68	1	63	5	253
Agrammus agrammus					1	176			1	176
Lateolabrax japonicus	74	6416	61	6846	23	2741	16	1030	174	17,033
Amblychaeturi-chthys hexanema	2	61	2	296			1	5	5	362
Oplegnathus fasciatus	1	110	1	220					2	330
Ditrema temminckii	7	257	4	174	1	88	8	616	20	1135
Saurida undosquamis	1	202	2	828	3	1010			6	2040
Engraulis japonicus	2	11	2	10			2	17	6	38
Pseudopleuron-ectes yokohamae	1	37							1	37
Sardinella zunasi	30	555	16	402	8	175	6	107	60	1239
Argyrosomus argentatus	16	2324	10	1695	9	1083	9	1569	44	6671
Sebastes schlegeli			1	38					1	38
Conger myriaster							1	55	1	55
Scomberomorus niphonius	43	9190	22	5558	19	5872	11	4093	95	24,713
Parapercis sexfasciata			1	54					1	54
Glossanodon semifasciatus	1	54			1	47			2	101
Chelidonichthys spinosus	22	1169	22	1777	7	623	10	655	61	4224
Inimicus japonicus	1	158							1	158
Sphyraena japonica			1	18					1	18
Platycephalus indicus	20	2884	20	2911	10	2695	8	1841	58	10,331
Konosirus puntatus	1650	123,786	881	77,107	243	27,710	40	5651	2814	234,254
Leiognathus nuchalis	424	8729	182	4041	41	852	35	612	682	14,234
Hexagrammos otakii	2	269							2	269
Pagrus major					2	135			2	135
Larimichthys crocea	1	22							1	22
Thryssa kammalensis	74	1006	36	491	22	361	11	144	143	2002
Aplysia kurodai							1	266	1	266
Sepiida					1	76			1	76
Octopus minor					1	210			1	210
Loligo japonica			1	61					1	61
Amphioctopus fangsiao	5	203	1	40			1	18	7	261
Oratosquilla oratoria	29	694	12	297	9	186	17	624	67	1801
Portunus trituberculatus	5	1050	3	276	4	596	5	1099	17	3021
Portunus trituberculatus	7	350	14	771	8	395	12	700	41	2216
Portunus sanguinolentus	5	474	1	123			5	481	11	1078
Marsupenaeus japonicus	4	123	3	73	6	199	4	115	17	510
Metapenaeus joyneri	1	8			1	9			2	17
Fenneropenaeus chinensis					1	38			1	38
Tonna luteostoma					1	197			1	197
Batillus cornutus	1	95							1	95
Total	591	47,768	445	43,172	272	33,251	96	10,340	1404	134,531

), 591 fish were caught with 50.5 mm, 445 fish were caught with 55.1 mm, 272 fish were caught with 60.6 mm, and 96 fish were caught with 67.3 mm mesh sizes. Clearly, the catch increased along with net size. The number of species caught was the most diverse, ranging from 30 species with a 50.5 mm mesh size, followed by 27 with 55.1 mm, 26 with 60.6 mm, and 22 with 67.3 mm mesh sizes (

Table 5. Catch by species and test gear mesh size.

Mesh Size (mm)	Number of Sampled Species	Weight (g)	Total Length (cm)	Fork Length (cm)	Bycatch Ratio (%)
50.5	30	75.0	19.7	17.1	23.2
55.1	27	87.5	20.5	17.9	26.4
60.6	26	114.0	22.0	19.4	40.6
67.3	22	142.3	23.5	20.7	71.6
Average		83.2	20.2	17.6	40.5

). On the other hand, the non-target species bycatch rate was highest at 71.6% with a 67.3 mm mesh size, followed by 40.6% with 60.6 mm, 26.4% with 55.1 mm, and 23.2% with 50.5 mm mesh sizes (Table 5). Slipmouth (Leiognathus nuchalis) accounted for 43.8% of the bycatch species regardless of mesh size. Although the amount of bycatch species varied by mesh size, Sea bass (Lateolabrax japonicus), Chefoo thryssa (Thryssa kammalensis), Japanese Spanish Mackerel (Scomberomorus niphonius), and Mantis shrimp (Squilla oratoria) were the most commonly caught in this order (Table 4).

3.2. Body Length Distribution of the Gizzard Shad

The total length of the gizzard shad ranged from 15.4 to 27.4 cm, averaging 20.2 cm, and the mode was 1146 for body length between 20 and 21 cm (Figure 5a). The captured gizzard shad for the fork length ranged from 13.4 to 24.2 cm, averaging 17.6 cm, and the mode was 933 for the total lengths ranging from 17 to 18 cm (Figure 5b). Based on this mode value, the larger individual distribution was higher than that of smaller individuals. The gizzard shads’ total length composition relative to the mesh size is presented in Figure 6a. As the mesh size increased from 50.5 to 55.1, 60.6, and 67.3 mm, the fish total length mode increased from 20 to 23 cm, and the fish total length composition also shifted toward larger total lengths. The gizzard shad fork length composition by mesh size is displayed in Figure 6b. As the mesh size increased from 50.5 to 55.1, 60.6, and 67.3 mm, the total length mode increased from 17 to 20 cm, and the fork length composition also shifted toward larger fork lengths. These results show the selectivity of size in catches relative to mesh size.

3.3. Mesh Selectivity Curve Estimation

Figure 7 presents the estimated mesh selectivity curve for each mesh size using the data displayed in Figure 6. The mesh selectivity curve equation estimated from these results is shown in Equation (1), where l is the gizzard shad total length, m is the mesh size, R is the total length-to-mesh size ratio, and $s\left(R\right)$ is the reference curve equation of the total length-to-mesh size ratio.

$$s\left(R\right)=s\left(\frac{l}{m}\right)=\exp \{\left(1.58R^3-23.11R^2+108.35R-159.86\right)-5.11\left)\right\}$$

(1)

As shown in Figure 7, the total length with a 0.5 gillnet selection ratio was estimated as 179.3, 195.6, 215.1, and 238.9 mm for 50.5, 55.1, 60.6, and 67.3 mm mesh sizes, respectively.

3.4. Statistics and Decision Tree Analysis Results

The descriptive statistics of variables used for machine learning analysis are displayed in

Table 6. Average variable distribution relative to mesh size.

Mesh (mm)	Mature	Score	CPUE	Bycatch (%)
50.5	0.04	0.74	11.46	20.31
55.1	0.00	0.75	5.45	21.61
60.6	0.00	1.50	2.38	27.24
67.3	0.26	1.58	0.44	39.74

. The bycatch and score revealed that the average value increased as the mesh size increased. Conversely, the average CPUE value increased as the mesh size decreased. As a result of analyzing the correlation between variables (Table 6), score and bycatch correlated positively with a 0.79 correlation coefficient. Concerning CPUE, a strong negative correlation was found at −0.73 (Figure 8).

In Model 1, the analysis results using decision trees among machine learning techniques were as follows (Figure 9 and Figure 10). Mesh size standards were classified as 0 (very good), 1 (good), 2 (bad), and 3 (very bad) (Model 1). In Model 1, when classifying the group with a score of 0 (very good), it was found that mesh size was classified with an approximate 60% accuracy with mesh sizes 57.85 mm or less (Figure 9). Groups 1 (good) and 2 (bad) were classified at a 50% rate for mesh sizes below 63.95 mm and above 57.85 mm (Figure 9). This model’s training and testing performance was about 58%.

In Model 2, when CPUE exceeded, bycatch and mature variables when classified score groups as 0 (very good), 1 (good), 2 (bad), or 3 (very bad) (Model 2). When CPUE exceeded 6.479, group 0 was classified with 98% accuracy. For Group 1 (good), when CPUE was 6.479 or less and bycatch was 26.228% or less, the group was classified with an 82% accuracy. When the CPUE was less than 6.479, bycatch exceeded 26.228%, and mature was less than 0.5, classifying it into Group 2 (bad). Moreover, this model exhibited high training and test scores of approximately 90%. Regarding feature importance, CPUE was 80.84%, bycatch was 16.16%, and mature was 0% in this model.

4. Discussion

This study conducted eleven test operations at sea on shad gillnets with different mesh sizes. Based on the data, mesh selectivity curves and decision tree analysis techniques were applied to derive the optimal mesh size. The mesh selectivity curve shifted to the right as the mesh size increased, meaning that as the mesh size increased, the body length of the fish caught also increased.

Selecting a size that catches individuals longer than the mature body size is one of the most essential aspects when considering the appropriate net size for fishing to allow the reproduction of the species and avoid the collapse of the exploited fish population. In this study, the body length (total length) distribution of the gizzard shad caught through eleven tests ranged from 15.4 to 27.4 mm (Figure 5a). Kim and Lee [21] reported that gizzard shads reach group maturity at a total length between 18 and 19 cm. Furthermore, in a study by NFRDI [37], the 50% group maturity size of gizzard shads was estimated as 15.8 cm based on the fork length, reaching group maturity at 17.9–18.1 cm. Based on these studies, the proportion of caught individuals with a body length less than 18 cm and a fork length less than 15.8 cm was 84 (approximately 3%) out of 2814 individuals, which is a small proportion. Moreover, all mesh sizes exceeded the mature length range in the 50% selection area. However, most fish that did not reach maturity were caught with a 50.5 mm mesh size (81 individuals).

Concerning fishery resource protection, most studies estimate the selectivity curve using body length information (maturity level) of fished individuals and then derive an appropriate mesh size from this [5,7,8,10]. Conversely, Kim et al. [38] derived an appropriate mesh size by considering economic aspects rather than mature individuals. However, although the appropriate mesh was estimated through CPUE and selectivity curves, a model was not derived by considering these variables in complex. Therefore, in this study, the score was calculated considering ecological (bycatch, maturity length) and economic perspectives (CPUE). In addition, we derived an appropriate mesh by creating a model that set standards for mesh and variables to classify groups with this score.

In Model 1, Group 0 (very good) was determined if the mesh was less than 52.8 cm. On the other hand, if it exceeded 63.95 cm, it was classified into Group 2 (bad). However, the model performance was low at approximately 58%, which may be the reason for the lack of individual information on shads at 60.6 and 67.3 mm due to the imbalance in the number of fish caught between mesh groups. Comparatively, Model 2 indicated a high performance of over 90%, and CPUE was determined as the most important criterion for group classification relative to feature importance (Figure 10). However, unlike the previous selectivity curve, the maturity and bycatch rate did not have a significant effect. CPUE was the most essential criterion for classifying groups because there was a severe catch imbalance between the number of test fishing operations.

CPUE was significantly higher than that reported for other meshes, with an average CPUE of 11.46 recorded with the 50.5 mm mesh size. During fishing, the catch amount differed by variables (environment, operator skill level), but the ecological characteristics of the target fish species also exhibited a significant impact. In particular, the gizzard shad is a species of the herring family and swims in groups of similar body size [8]. Therefore, many fish are caught if the fish size appropriately matches that of the mesh size [8]. Conversely, the catch is significantly lower if the mesh size is larger than that of the fish [8]. Thus, the catch amount widely varies depending on the mesh size.

Park et al. [5] used an encircling gillnet with a different fishing method than the shad gillnet used in this study, but the optimal mesh size was estimated to be between 50 mm and 60 mm. Therefore, when deriving an appropriate mesh for gizzard shad fishing by applying two analysis methods, using a mesh size of 57.85 mm or less would be appropriate for the sustainable exploitation of gizzard shad under the current economic perspective. Notably, the mesh sizes of 50.5 and 55.1 mm used in the present study produced appropriate results for the sustainable exploitation of the gizzard shad in South Korea.

5. Conclusions

This study developed a selectivity curve estimation and machine learning model using catch data to derive an appropriate mesh size for the shad gillnet. The estimation of selectivity curves revealed that all mesh sizes exceeded the mature length range within the 50% selection area. Furthermore, utilizing machine learning models to derive the optimal mesh size through combinations of various variables resulted in identifying the mesh size that performed best under conditions of 57.85 mm or less. Therefore, based on these two outcomes, the appropriate mesh size can be concluded to be within the range of 50.5 to 55.1 mm, using the mesh size employed in this experiment as a reference. However, although data were measured at sea eleven times, due to the ecological nature of gizzard shad as they move in groups, there was a significant deviation in the catch data between fishing days. This model can be improved in future research by conducting additional sea experiments within the currently derived range. As a result of this, converging the range estimated for the current section into a point estimate will be possible.