Predicting Skipjack Tuna Fishing Grounds in the Western and Central Paciﬁc Ocean Based on High-Spatial-Temporal-Resolution Satellite Data

: Skipjack tuna are the most abundant commercial species in Taiwan’s pelagic purse seine ﬁsheries. However, the rapidly changing marine environment increases the challenge of locating target ﬁsh in the vast ocean. The aim of this study was to identify the potential ﬁshing grounds of skipjack tuna in the Western and Central Paciﬁc Ocean (WCPO). The ﬁshing grounds of skipjack tuna were simulated using the habitat suitability index (HSI) on the basis of global ﬁshing activities and remote sensing data from 2012 to 2015. The selected environmental factors included sea surface temperature and front, sea surface height, sea surface salinity, mixed layer depth, chlorophyll a concentration, and ﬁnite-size Lyapunov exponents. The ﬁnal input factors were selected according to their percentage contribution to the total efforts. Overall, 68.3% of global datasets and 35.7% of Taiwanese logbooks’ ﬁshing spots were recorded within 5 km of suitable habitat in the daily ﬁeld. Moreover, 94.9% and 79.6% of global and Taiwan data, respectively, were identiﬁed within 50 km of suitable habitat. Our results showed that the model performed well in ﬁtting daily forecast and actual ﬁshing position data. Further, results from this study could beneﬁt habitat monitoring and contribute to managing sustainable ﬁsheries for skipjack tuna by providing wide spatial coverage information on habitat variation.


Introduction
Skipjack tuna (Katsuwonus pelamis) is known for being highly migratory and is widely distributed in the Western and Central Pacific Ocean (WCPO) [1]. The lack of a swim bladder allows skipjack tuna to flexibly move near the surface of water [2], and they can be caught by industry fishing equipment, such as purse seine nets and pole and line [3,4]. As an economically harvested species, skipjack tuna are mainly sold to canneries [5]. This fishery harvest ranks third among the most fished species globally, and catches have more than doubled since the 1980s [6]. The increasing catches of skipjack tuna are evident due to the flourishing large-scale, international purse seine fisheries that predominate in the WCPO [3,7].
Several studies have noted that environmental parameters influence skipjack tuna distribution selection, resulting in apparent spatial shifts in the skipjack tuna [8]. The physiology of skipjack also plays a role: They avoid cooler water and inhabit warm water than CPUE [29,30]. Therefore, we used the location of fishing efforts as an indicator to establish our HSI model.
The objectives of this study were to investigate the connection of skipjack tuna habitat suitability with environmental factors in the WCPO using fishery data derived from automatic identification system (AIS) and satellite-derived oceanic variables. The results will allow us to better understand the fine scale dynamics of skipjack tuna. Furthermore, this study provides fundamental information on expressing relationships in ecological systems, which is essential in the sustainable development of fishery resources.

Skipjack Tuna Fishery Data
The traditional purse seine fishing areas of Taiwan are mainly distributed in the WCPO spanning from 130 • E-150 • W to 20 • N-20 • S, and all data were collected in this region ( Figure 1). Two types of data were collected: Global datasets and Taiwan's logbook data. Global datasets were obtained from Global Fishing Watch, which analyzed AIS data and identified fishing activities during the period from 2012 to 2016 [31]. The datasets contain fishing activities at 0.01 • resolution, and vessel presence by flag state and fishing gear type Taiwanese data were gathered from the Overseas Fishery Development Council (OFDC) of Taiwan during the year of 2016. Logbook data comprised daily fishing positions (longitude and latitude in 0.01 • Spatial grid), the school type of fish, the set type of fishing, and the total catch (in tons). To facilitate the compilation of all data from inconsistent spatial scales, the original data were aggregated into the same grid resolution as environmental data.
Remote Sens. 2021, 13, x FOR PEER REVIEW 3 of 17 attempted to show that fishing effort may be a more reliable indicator than CPUE [29,30]. Therefore, we used the location of fishing efforts as an indicator to establish our HSI model. The objectives of this study were to investigate the connection of skipjack tuna habitat suitability with environmental factors in the WCPO using fishery data derived from automatic identification system (AIS) and satellite-derived oceanic variables. The results will allow us to better understand the fine scale dynamics of skipjack tuna. Furthermore, this study provides fundamental information on expressing relationships in ecological systems, which is essential in the sustainable development of fishery resources.

Skipjack Tuna Fishery Data
The traditional purse seine fishing areas of Taiwan are mainly distributed in the WCPO spanning from 130° E-150° W to 20° N-20° S, and all data were collected in this region ( Figure 1). Two types of data were collected: Global datasets and Taiwan's logbook data. Global datasets were obtained from Global Fishing Watch, which analyzed AIS data and identified fishing activities during the period from 2012 to 2016 [31]. The datasets contain fishing activities at 0.01° resolution, and vessel presence by flag state and fishing gear type Taiwanese data were gathered from the Overseas Fishery Development Council (OFDC) of Taiwan during the year of 2016. Logbook data comprised daily fishing positions (longitude and latitude in 0.01° spatial grid), the school type of fish, the set type of fishing, and the total catch (in tons). To facilitate the compilation of all data from inconsistent spatial scales, the original data were aggregated into the same grid resolution as environmental data. Industrial tropical purse seiners deploy fish aggregating devices (FADs), which are known to provide geo-locations of objects to attract the target tuna [32,33]. Some FADs are powerful in not only locating valuable targets, but also highlight the signals of Industrial tropical purse seiners deploy fish aggregating devices (FADs), which are known to provide geo-locations of objects to attract the target tuna [32,33]. Some FADs are powerful in not only locating valuable targets, but also highlight the signals of nontargets [34]. Hence, we used only free-swimming schools from Taiwan's logbook data for further analysis.

Remotely-Sensed Environmental Data
The oceanographic variables (Table 1) used for modelling the habitat suitability index include sea surface temperature (SST) and its gradient (SST front; calculated as the peak and ridge temperature features in the same pixel) [35], sea surface salinity (SSS), sea surface height (SSH), mixed layer depth (MLD), chlorophyll a concentration (CHLA), and finite-size Lyapunov exponents (FSLE). Oceanic fronts are known for their characteristic of increased abundance and diversity of taxa ranging from phytoplankton to top predators [36]. The SST front detection algorithm is based on the gradient magnitude, which reveals the fine scale of boundaries between water masses ( Figure 2). These operational variables were chosen because they are known to affect catches of tuna [2,19,37,38]. operation, the upper level (0 to 5 m) oceanographic factors were selected for model construction.
Daily mean FSLE data were download from Archiving Validation and Interpretation of Satellite Oceanographic (AVISO) data with a spatial resolution of 1/25 degree (~4 km) (available at https://www.aviso.altimetry.fr/ (accessed on 17 January 2021)). All datasets and SST frontal detections were decoded and integrated by code programming in Interactive Data Language (IDL 8.7.2) using the Advanced Math and Stats Module (IMSL). All environmental variables were resampled to 0.08-degree (~8 km) resolution on the basis of the coarsest scale.

Habitat Suitability Index Model
We started by fitting global datasets from 2012 to 2015 with remote sensing data to establish the HSI model. First, the suitability index (SI) was calculated using the frequency distribution of environmental parameters. The SI was described as a score, ranking from 0 to 1 for inappropriate and optimal habitats, respectively: This is an example of the equation: (1) For the end user, the Copernicus Marine Environment Monitoring Service (CMEMS website provides level-4 products of daily mean SST, SSS, SSH, and MLD that were generated using a statistical model based on both satellite-derived data and in situ observations. Remote Sens. 2021, 13, 861 5 of 16 CHLA concentration was derived as a weekly mean because of low coverage of cloudfree pixels in daily satellite-derived images (available at http://marine.copernicus.eu/ (accessed on 17 January 2021)). For the SST product, multiple satellites are integrated from the Advanced Very-High-Resolution Radiometer (AVHRR), Environmental Satellite (Envisat), Aqua, and the Tropical Rainfall Measuring Mission (TRMM). The SSS product is based on the data derived from the satellite Soil Moisture Ocean Salinity (SMOS) and in situ salinity measurements. SSH and MLD products are processed from all altimeter missions: Jason-3, Sentinel-3A, HaiYang Daily mean FSLE data were download from Archiving Validation and Interpretation of Satellite Oceanographic (AVISO) data with a spatial resolution of 1/25 degree (~4 km) (available at https://www.aviso.altimetry.fr/ (accessed on 17 January 2021)). All datasets and SST frontal detections were decoded and integrated by code programming in Interactive Data Language (IDL 8.7.2) using the Advanced Math and Stats Module (IMSL). All environmental variables were resampled to 0.08-degree (~8 km) resolution on the basis of the coarsest scale.

Habitat Suitability Index Model
We started by fitting global datasets from 2012 to 2015 with remote sensing data to establish the HSI model. First, the suitability index (SI) was calculated using the frequency distribution of environmental parameters. The SI was described as a score, ranking from 0 to 1 for inappropriate and optimal habitats, respectively: This is an example of the equation: In Equation (1), α and β are the regression coefficients, which were modified by the least-squares method to assess the residuals between observations and functions of SI [21]. env is the value for each environmental parameter. After calculating the SI values, the SIs were subsequently developed to determine the overall habitat preference. We used the geometric mean model (GMM) to set up the HSI model [21,39].
The GMM model is described as follows: where n is the number of selected environmental parameters for setting up the HSI algorithm, and SI i is the SI for the ith environmental variable. Similar to the SI, the HSI has values ranging from 0 to 1.0, indicating not suitable to optimal, respectively [26,39]. HSI values equal to or higher than 0.6 were regarded as indicating potential fishing grounds [25].

Calculation of Predicting Rate
To evaluate the model's performance, we calculated the habitat suitability index using environmental data in the year of 2016 and compared it with untrained fishing activities data, included in global datasets and Taiwan's logbook data over the same period, to estimate the fishing grounds of skipjack tuna in the WCPO. The distance between the observed fishing effort and the boundary of potential fishing grounds derived from the HSI model is the clearest approach to show the ability of our model to locate potential fishing grounds. The accuracy rates were defined by the ratio of fishing activity occurring within different distances to the potential fishing grounds [11]. The workflow of this study is demonstrated in (Figure 3). HSI values equal to or higher than 0.6 were regarded as indicating potential fishing grounds [25].

Calculation of Predicting Rate
To evaluate the model's performance, we calculated the habitat suitability index using environmental data in the year of 2016 and compared it with untrained fishing activities data, included in global datasets and Taiwan's logbook data over the same period, to estimate the fishing grounds of skipjack tuna in the WCPO.
The distance between the observed fishing effort and the boundary of potential fishing grounds derived from the HSI model is the clearest approach to show the ability of our model to locate potential fishing grounds. The accuracy rates were defined by the ratio of fishing activity occurring within different distances to the potential fishing grounds [11]. The workflow of this study is demonstrated in (Figure 3).

Variations of Skipjack Tuna in the WCPO
The quarterly spatial distribution of fishing efforts from 2012 to 2015 at 1 degree resolution ( Figure 4) showed quarterly variations. The global datasets are indicated in gray circles, which were widely distributed in the Pacific Ocean for all quarters. The Taiwanese vessels are shown by red dots and were mainly concentrated in the exclusive economic zone (EEZ) of Papua New Guinea, Nauru, Federated States of Micronesia, and west Kiribati (165° E-175° W and 5° N-5° S) and extended to Central Kiribati (170° W) in the second quarter. In 2016 ( Figure 5), location variations were clearly seen in the third quarter. Global datasets showed less activity in the east high seas (170° W-160° W). Meanwhile, Taiwanese vessels ( Figure 6) were concentrated in the EEZ of Nauru and Kiribati (165° E-180° E). Due to the lack of logistic lines and supplemental ports, Taiwanese purse seine vessels did not tend to fish on the east side of the Pacific Ocean from the commercial profit perspective.  Meanwhile, Taiwanese vessels ( Figure 6) were concentrated in the EEZ of Nauru and Kiribati (165° E-180° E). Due to the lack of logistic lines and supplemental ports, Taiwanese purse seine vessels did not tend to fish on the east side of the Pacific Ocean from the commercial profit perspective.

Suitability Index Analysis and Habitat Suitability Index Model
The intention for model development based on environmental variables and fishery data was to classify the preferred regions for skipjack tuna from a remote sensing perspective. A total of 737,167 datapoints were used in the HSI model, taken from the global datasets in the period of 2012-2015. The suitability index equations for each parameter, derived from global datasets, are shown in Table 2. All environmental parameters were analyzed for statistical significance. For purse seiners, the optimal range (SI 0.6) of skipjack tuna was defined as a slightly hotter SST, higher than 29.6 • C (Figure 7a (Figure 7g). Furthermore, the optimal ranges of SST, SST front, SSH, SSS, and FSLE accounted for 60.00%, 62.66%, 51.26%, 57.40%, and 64.50% of total efforts, respectively ( Table 2). distributed in Nauru and Kiribati waters, where the adjacent water masses on the eas side are cold and on the west side are warm. The fishing spots were in warm wate where the SST was higher than 29 ℃ and there was a low gradient of SST front (Figure 8 left column). Furthermore, the high-gradient SST front was more associated with fishing grounds. The strong SST front did not specifically appear in the fishing datasets nevertheless, the fishing grounds were not far from strong fronts (Figure 8, righ column).  The daily distributions of operation locations from the 20 to 23 August 2016, with SST and SST front images are shown in Figure 8. The fishing spots were mainly distributed in Nauru and Kiribati waters, where the adjacent water masses on the east side are cold and on the west side are warm. The fishing spots were in warm water where the SST was higher than 29 • C and there was a low gradient of SST front (Figure 8, left column). Furthermore, the high-gradient SST front was more associated with fishing grounds. The strong SST front did not specifically appear in the fishing datasets; nevertheless, the fishing grounds were not far from strong fronts (Figure 8, right column).

Accuracy of the HSI Model
Five suitable ranges of environmental variables, which accounted for up to 50% of total efforts, were used, i.e., SST, SST front, SSH, SSS, and FSLE (Table 2). These were selected to derive the HSI model for skipjack tuna. The period of 2016 was chosen to calculate the accuracy of the HSI model by comparing data with fishery efforts, including global datasets and Taiwanese logbook data. In the overview of 2016, 68.3% of global data was within 5 km of suitable areas, whereas 94.9% were within 50 km (Figure 9a). In contrast, 35.7% of Taiwanese logbook data occurred within 5 km, and 79.6% were found within 50 km of suitable areas (Figure 9c). The accuracy of the model without SST fronts was compared, and found to decrease to 62.3% and 91.9% within 5 and 50 km, respectively, in global data (Figure 9b). Taiwanese logbook data showed increased accuracy (43.3%) in the area within 5 km but lower accuracy of 74.9% within 50 km (Figure 9d). As a consequence, SST fronts is the parameter which improves the predictability of fishing grounds in the suitability index-based models.

Accuracy of the HSI Model
Five suitable ranges of environmental variables, which accounted for up to 50% of total efforts, were used, i.e., SST, SST front, SSH, SSS, and FSLE (Table 2). These were selected to derive the HSI model for skipjack tuna. The period of 2016 was chosen to calculate the accuracy of the HSI model by comparing data with fishery efforts, including global datasets and Taiwanese logbook data. In the overview of 2016, 68.3% of global data was within 5 km of suitable areas, whereas 94.9% were within 50 km ( Figure  9a). In contrast, 35.7% of Taiwanese logbook data occurred within 5 km, and 79.6% were found within 50 km of suitable areas (Figure 9c). The accuracy of the model without SST fronts was compared, and found to decrease to 62.3% and 91.9% within 5 and 50 km, respectively, in global data (Figure 9b). Taiwanese logbook data showed increased

The Outputs of the HSI Model and Purse Seine Fishing
The outputs of the habitat model in 2016 are shown in Figure 5, and global datasets showed that fishing activities generally occurred in the equatorial region, except for the Phoenix Islands Protected Area (175 • W-170 • W). The fishing hours decreased in the high seas (170 • W-160 • W) from quarter 1 and were nearly absent in quarter 3. Notable, in the Central Pacific, Nauru and West Kiribati are two important fishing grounds for purse seiners, showing long fishing hours throughout the year. The habitat derived from the HSI model shifted northwest to the Federated States of Micronesia and Papua New Guinea in quarter 2, and continued to quarter 4. Simultaneously, the length of fishing hours showed the same trend, moving northwest and extending to the south of the EEZ in the Federated States of Micronesia. In general, the suitable habitat displayed a large horizontal shift in the period of quarters 1 and 2, but was relatively stable in quarter 2-quarter 4. In comparison, Taiwanese purse seine vessels showed less variations in the vertical and horizontal directions ( Figure 6). In quarter 1, fishing efforts mainly occurred in the EEZ of Nauru, the junction (4 • S) of west Kiribati and Tuvalu, Solomon Islands. The high seas (178 • W-176 • W) between Tuvalu and central Kiribati also showed dense fishing activities. In quarter 2, fishing locations shifted northward and westward, similar to the shift in suitable habitat. Skipjack tuna were caught across the EEZ of Papua New Guinea and along the line from Federated States of Micronesia to Marshall Islands. In quarter 3, fewer suitable habitats were detected in Tuvalu, Tokelau, and Cook Islands, whereas the Western Pacific was highlighted as a suitable fishing ground. The catches were also mainly concentrated in two areas: the central area between two high seas (151 • E-158 • E) and the EEZ of Nauru to Central Kiribati. Suitable habitats were no longer present in the Western Pacific in quarter 4. Instead, the HSI model marked the south edge of the EEZ of Nauru, the north of the Federated States of Micronesia, and all of Tokelau, as suitable habitats. Catches were concentrated on the south side of Nauru and north side of Papua New Guinea, and were sporadic between Pohnpei and Majuro.
Remote Sens. 2021, 13, x FOR PEER REVIEW 11 of 17 accuracy (43.3%) in the area within 5 km but lower accuracy of 74.9% within 50 km (Figure 9d). As a consequence, SST fronts is the parameter which improves the predictability of fishing grounds in the suitability index-based models.

The Outputs of the HSI Model and Purse Seine Fishing
The outputs of the habitat model in 2016 are shown in Figure 5, and global datasets showed that fishing activities generally occurred in the equatorial region, except for the Phoenix Islands Protected Area (175° W-170° W). The fishing hours decreased in the high seas (170° W-160° W) from quarter 1 and were nearly absent in quarter 3. Notable, in the Central Pacific, Nauru and West Kiribati are two important fishing grounds for purse seiners, showing long fishing hours throughout the year. The habitat derived from the HSI model shifted northwest to the Federated States of Micronesia and Papua New Guinea in quarter 2, and continued to quarter 4. Simultaneously, the length of fishing hours showed the same trend, moving northwest and extending to the south of the EEZ in the Federated States of Micronesia. In general, the suitable habitat displayed a large horizontal shift in the period of quarters 1 and 2, but was relatively stable in quarter 2quarter 4. In comparison, Taiwanese purse seine vessels showed less variations in the vertical and horizontal directions ( Figure 6). In quarter 1, fishing efforts mainly occurred in the EEZ of Nauru, the junction (4° S) of west Kiribati and Tuvalu, Solomon Islands. The high seas (178° W-176° W) between Tuvalu and central Kiribati also showed dense fishing activities. In quarter 2, fishing locations shifted northward and westward, similar to the shift in suitable habitat. Skipjack tuna were caught across the EEZ of Papua New Guinea and along the line from Federated States of Micronesia to Marshall Islands. In quarter 3, fewer suitable habitats were detected in Tuvalu, Tokelau, and Cook Islands, whereas the Western Pacific was highlighted as a suitable fishing ground. The catches Overall, in both datasets, the EEZs of Nauru and Kiribati were the most popular fishing areas (163 • E-180 • E, 2 • N-3 • S), except in quarter 2. The east high seas surrounding Kiribati (167 • W-160 • W) were marked as a suitable habitat in quarters 1 and 2, and the fishing activities in the global dataset presented the same spatial distribution for the same period, but Taiwanese vessels were absent in this region. Taiwanese vessels appeared to avoid crossing the vertical line of 180 • E, showing few records in the narrow high seas. An increase in fishing activities was also found in both datasets: the recorded number fishing activities in both datasets in the study area increased by 1.5 times from quarter 1 to 4. Taiwanese and global datasets recorded the highest fishing activities in quarters 3 and 4, respectively.

Discussion
The HSI model theory led us to infer characteristics of the skipjack tuna habitat in the WCPO and build a forecasting system using fishery and satellite remote sensing data to locate potential fishing grounds. The fishery data collected were not field survey quality because fishing masters determined fishing locations according to their own choice and not by random selection. However, fishery data are easy to acquire, and low-cost species survey data sets are available to scientists [2]. Here we obtained fishing effort data that covered a wide geographical area, and data were from all operational purse seine vessels in the WCPO. The advantage of using remote sensing data is the efficiency in measuring parameters on the oceanic scale compared to on-site investigations [40]. By programming code, we were able to construct an automatic procedure to locate interesting areas for finding target fish in just a few minutes.
The global purse seine fishery industry operates in the Pacific Ocean over a wide distribution from 140 • E to 150 • W. In contrast, traditional fishing grounds for Taiwanese vessels mainly occurred around the Federated States of Micronesia, Papua New Guinea, Solomon Islands, and Kiribati from 1997 [41]. Skipjack tuna follow a brief route to the south for spawning and retreat to the north during the summer [42]. Fishing efforts extended south to the Solomon Islands in quarters 1 and 4, and they withdrew from the south in quarters 2 and 3 in both datasets (Figures 4-6). These shifts seem to reflect the migration route of skipjack tuna following the south-north pathway. The high sea pocket was enclosed from 1 January 2010 for conservation management [43]. The footprints of Taiwanese purse seines were present around the high sea pocket (i.e., they were fishingthe-line), and the fishing masters obtained benefits from the spillover of the enclosed area ( Figure 4).
Typically, the western Pacific is a warm pool with an annual maximum sea surface temperature of 30 • C, and a higher density of skipjack tuna gather there [22,23]. The SST in this study was slightly higher compared to previous studies, ranging from 20.5 to 26.0 • C for skipjack in the Western North Pacific Ocean [2] and over 24 • C in the Western Pacific Ocean [1]. A possible explanation for this might be the sampling bias derived from the tendency for operational fleets to find targets in higher SSTs. Another possible explanation is a regional bias derived from the different resource types of data and collecting areas. There are, however, other possible explanations. It is thought that predators (tuna) cluster around the front area where the prey is located [44], and thermal fronts are also associated with a high probability of the skipjack population [17].
The findings of the SST front were confirmed in the lower gradient change from 0.01 to 0.11 • C/km. The most likely cause of this lower gradient change is that fishing spots are located at the junction of the convergence zone of the warm pool and the tongue of cold surface water [23]. The cold tongue is nutrient-enriched, which benefits prey, and the warm pool offers a place for skipjack tuna to cluster. Skipjack tuna move to cooler water for ingestion and swim into warm water due to its comfortable temperature. The fishing activities of skipjack tuna occur in warm water, near the convergence zone of warm and cold water where there is less gradient change compared to the junction of two water masses. Our findings indicate that the actual fishing grounds were closely associated with adjacent areas of strong SST fronts rather than in the center of strong fronts (Figure 8).
A higher value of SSH influences the suitable range of skipjack tuna, and this was confirmed in our study. For example, SSH ranged from 0 to 50 cm in the Western North Pacific Ocean [2], and 70-100 cm in the WCPO [14,15], which were defined by the generalized additive model (GAM) method, and 80-90 cm by the empirical cumulative distribution function in Sri Lankan waters [13]. The intermediate values of these findings coincided with our results ranging from 60 to 71 cm (Figure 7c), given that the datasets in our study had broad coverage. The salinity broadly ranges from 33.0 to 37.2 PSU for the world oceans [16]; 5th and 95th percentile values ranged from 34.9 to 35.8 PSU in the Southwestern Atlantic according to the GAM method [17], and 30.3 to 36.2 PSU was recorded in the Eastern central Atlantic and Western Indian Oceans by the ecological niche model [10]. High salinity values above 35.8 PSU present oligotrophic conditions and lower primary productivity, which cannot support feeding requirements for skipjack tuna in tropical waters [17]. A suitable range from 34.29 to 35.28 PSU (Figure 7d) was seen under the border limit and is consistent with previous studies.
Skipjack tuna do not resist hypoxic water where there is less than 3.5 mL O 2 per liter [16], and this restricts skipjack tuna to inhabit the mixed layer above the thermocline [2]. The MLD from 6 to 158 m was confirmed as a favorable feeding habitat for skipjack [11]. A fishing master works towards better sinking performance, where faster and greater sinking depth means a higher probability of a successful set [45]. Underwater nets are attached to FADs; the length can reach a depth of 50 m [46], and the common average depth is 15-20 m [33]. Moreover, MLD associated with oxygen restriction affects the vertical distribution of tuna, and larger skipjacks with greater requirements are more spatially confined [9]. The suitable range of MLD reflects the selection of fishing operations more than the suitability of the skipjack tuna habitat. In our findings, the suitable range of MLD showed a highly left-skewed distribution (Figure 7e) and accounted for 20% of total efforts ( Table 2). We propose that the MLD, which implies the limit of oxygen concentration, does not directly reflect fishing activities.
Chlorophyll concentration has direct or indirect relationships with the foraging distribution, and affects the predators [2]. The specific value of 0.2 mg/m 3 was described as an indicator for the highest CPUE and fishing frequency in the gulf of the Bone-Flores Sea, Indonesia [12]. A similar range of 0.1 to 0.3 mg/m 3 was also shown in the Western North Pacific [2]. A widely defined favorable range between 0.13 and 5.27 mg/m 3 exists in the Atlantic and Indian Oceans [11]. Furthermore, fishing positions in the aforementioned studies were mostly located in coastal areas known for high nutrient concentrations. There is a discrepancy between our findings ranging from 0.03 to 0.04 mg/m 3 ( Figure 7f) and a previous study on the suitable range of CHLA. We assumed this difference resulted from feeding habits. As a piscivorous feeder, skipjack tuna also depends on various prey [47]. Skipjack tuna can easily adapt their survival strategy to the local prey composition [47,48], different food consumptions of tuna between FAD-associated and -unassociated schools [49], and feeding habit variations in different current systems [47]. Thus, the feeding strategy in distant waters with poorer nutritional conditions seems to differ from the relatively high CHLA region.
Unlike other physical oceanic parameters, the FSLE is new dynamic concept developed using the Lagrangian technique. This new concept crucially influences the marine top predators, indicating their foraging behavior, movement distribution, etc. [18][19][20]. For example, strong Lagrangian fronts imply a zooplankton concentration and larger quantities of fish, and more prey, in the Southern Indian Ocean [18]. Tuna fishermen track strong Lagrangian coherent structures where three times more profits per trip are expected in the U.S. in California [20]. A strong FSLE represents an area of aggregated fish, but this strong oceanic dynamic also represents stirring and strongly organizes the fluid motion. The prerequisites for successful purse seine fishing operations are the direction and speed of the wind and current [50], which are strong under current due to the shear force, and this may cause damage to the net. FSLE ranges close to zero (Figure 7g) indicate better conditions for purse seine fishing, although they may not represent the highest density of skipjack tuna.
After determining the optimal range for each environmental parameter relevant to skipjack tuna from global datasets, the accuracy of the HSI model was determined (Figure 7). HSI values in the accumulation map were appreciably concurrent with fishing spots for 2016 ( Figure 5). The present model appears to have defects in describing the spatial distribution in the first quarter ( Figures 5 and 6); the HSI values were highly accumulated in Tokelau and Cook Islands waters. Meanwhile, the actual fishing spots were mainly distributed in the region on the west side of the Pacific Ocean, for either global or Taiwanese data. The major distant water fishing nations, which include Korea (15%), Japan (14%), Taiwan (12%), and the USA (11%), accounted for all of the WCPO tuna catches by weight [51]. This list of major nations mostly comprises Asian nations. Considering multiple essential factors, such as increased supply line costs and oil prices, it is not difficult to imagine that Asian nations tend to fish on the western side of the Pacific. Even if most of the fish were on the east side of the Pacific, fishing vessels still fish in adjacent areas rather than chasing distant targets. Furthermore, fishing is moving toward FAD-based strategies instead of free-swimming school sets. In addition, the advantages of FADs are shown in three important aspects: the high success rates, stable schooling situations, and higher average catch [52]. Distinguishable behaviors of skipjack tuna include moving over long distances or their association with floating objects in the El Niño-Southern Oscillation (ENSO) [53]. Purse seine vessels no longer chase free-swimming schools, and powerful multifunction buoys are the first choice. Most FAD sets are done in the early morning, before sunrise, and most can still search for free-swimming schools during the day [46]. Therefore, it can be concluded that the FAD and the associated cost showed changes in fishing activity patterns and moving strategies.
In Taiwanese logbook data, about 20% of free-swimming sets were far from the optimal habitat boundary (>50 km), whereas the global data represented 5% of all sets, suggesting that the model performed well in fitting daily forecasting habitats and actual fishing position. However, quarterly accumulation of HSI appears to coincide with defects in quarter 1 ( Figure 5).

Conclusions
This study successfully developed an HSI model, which could be an efficient tool to forecast potential fishing grounds because it projects optimal habitat selection for skipjack tuna using fishery data and environmental parameters from a high spatial-temporal resolution and remote sensing perspective. Our automatic process could be ideally used to assist in management scenarios, to avoid marine protected areas, and to ensure proper areas are fished and eco-friendly regulations are followed. The dominant fishing grounds were not far from areas of strong SST fronts. Moreover, the model without SST fronts was compared with the optimal model to examine how SST fronts can improve the predictability of fishing grounds in the HSI model. Overall, the accuracies of the two datasets generally declined in the model without SST fronts. The decrease in SI elements (from 5 SI to 4 SI) implies a looser filter for determining suitable habitat. The model without SST fronts did not benefit from a loose filter to obtain higher accuracy compared with the optimal model. The result shows that SST fronts can improve the predictability of fishing grounds.
Future work should expand on our algorithm to improve the forecasting rate and provide precise directions of optimal fishing areas. In addition, considering more oceanic parameters, such as dissolved oxygen [9], could be helpful in detailing the habitat suitability index for skipjack tuna, in addition to identifying both fishing and weather conditions to determine optimal fishing conditions and fish locations.

Informed Consent Statement: Not applicable.
Data Availability Statement: GFW data used in this study are available from (www.globalfishingwatch. org (accessed on 17 January 2021)).