Recruitment forecasting of yellowfin tuna in the eastern Pacific Ocean with artificial neuronal networks

Ecological Informatics 36 (2016) 106–113

Contents lists available at ScienceDirect

Ecological Informatics journal homepage: www.elsevier.com/locate/ecolinf

Recruitment forecasting of yellowfin tuna in the eastern Pacific Ocean with artificial neuronal networks Laura Karen Torres-Faurrieta a, Michel J. Dreyfus-León b,⁎, David Rivas c a b c

Posgrado en Ecología Marina, CICESE, Ensenada, Baja California, Mexico Centro Regional de Investigación Acuícola y Pesquera Ensenada, Instituto Nacional de la Pesca, Ensenada, Baja California, Mexico Departamento de Oceanografía Biológica, CICESE, Ensenada, Baja California, Mexico

a r t i c l e

i n f o

Article history: Received 1 July 2016 Received in revised form 12 October 2016 Accepted 14 October 2016 Available online 18 October 2016 Keywords: Yellowfin tuna Recruitment NARX Artificial neural network Forecasting

a b s t r a c t The recruitment of yellowfin tuna in the eastern Pacific Ocean is modeled based on oceanographic as well as biological parameters, using two nonlinear autoregressive network models with exogenous inputs (NARX). In the first model (Model 1) the quarterly recruitment is modeled considering eastern Pacific global oceanographic conditions: the Southern Oscillation Index (SOI), the Pacific Decadal Oscillation (PDO), and spawners biomass. In Model 2, recruitment is predicted based on sea surface temperature, wind magnitude, and oceanic current magnitude of a smaller area within the eastern Pacific Ocean, considered as relevant for spawning and recruitment, and total spawners biomass. The correlation coefficient between the ANN recruitment estimate and the “real” recruitment is r N 0.80 in both models. Series of sensitivity analysis suggest that the SOI and the sea surface temperature are the most important variables for the recruitment in Model 1 and Model 2 also show that warm sea surface favors recruitment. A forecasting model under different climatological scenarios indicates that the recruitment of yellowfin tuna could be higher in the period 2015–2020 compared to the ones registered in the period 2009–2013. © 2016 Elsevier B.V. All rights reserved.

1. Introduction Yellowfin tuna (Thunnus albacares) is the second species in terms of its contribution to global tuna catches. The yellowfin tuna (YFT) fishery in the eastern Pacific Ocean (EPO) contributes yearly with N 200,000 tons (IATTC, 2015). YFT in the EPO is managed by a regional fishery organization, the Interamerican Tropical Tuna Commission, as a single stock. Spawning occurs year-round, extending roughly from 26°N to 14°S, and from the coast to 140°W (Shaefer, 1998). Although YFT tuna is a multiple spawner, recruitment of the young fish to the fishery is highly variable, with a seasonal component. It has also experienced different productivity regimes (IATTC, 2013). These regime variations have been associated with oceanographic conditions, especially with sea surface temperature and other parameters affecting the feeding dynamics of the larvae. For example, wind modulates the encounter rates of the larvae with its prey through turbulence (Haury et al., 1990; Kimura et al., 2004), as well as primary production through upwelling (Cury and Roy, 1989). Moreover, oceanic currents must have effects on the dispersion and transport of larvae to areas with different concentration of prey biomass.

⁎ Corresponding author at: 511 E. San Ysidro Blvd. 2430, San Ysidro, CA, USA. E-mail address: [email protected] (M.J. Dreyfus-León).

http://dx.doi.org/10.1016/j.ecoinf.2016.10.005 1574-9541/© 2016 Elsevier B.V. All rights reserved.

Recruitment (incorporation of juveniles to the fishing stock) is a very important variable as a biomass generator. Accurate prediction of recruitment to the fishery is an important tool for the management structure of any fish stock being exploited (Dreyfus-León and Schweigert, 2008), and it is useful with the aim of knowing the future status of a stock. However, there are difficulties in the yellowfin recruitment forecasting since no apparent relation between the amount of spawners and recruitment exists and it is quite possible that recruitment is shaped during early life stages, especially in the planktonic stage. Thus, the recruitment may be influenced by biotic and abiotic factors; knowing these possible relations is of paramount importance and such relations are generally highly non-linear. A very suitable option to explore non-linear relations between certain variable and other predictive variables is the implementation of Artificial Neural Networks (ANNs). ANNs are a powerful tool to explore complex, nonlinear biological problems (Chen and Ware, 1999; Yáñez et al., 2010). ANNs are computer algorithms that simulate in a strongly simplified way the activity of neurons and information processing in the human brain (Jarre-Teichmann et al., 1995). They create their own organization and representation of information they receive through a learning stage called training. ANNs are capable to generalize the system as a whole and respond appropriately to data or situations they have not been exposed previously (Lek and Guégan, 1999; Zhang et al., 1998).

L.K. Torres-Faurrieta et al. / Ecological Informatics 36 (2016) 106–113

The purpose of this paper is to present ANNs, specifically Nonlinear Autoregressive models with exogenous Inputs (NARX), to predict recruitment of YFT in the EPO from oceanographic parameters used as proxy of conditions influencing natural mortality at pre-recruit stages, and to identify variables that are more important shaping recruitment.

2. Methods 2.1. Overview YFT recruitment is modeled with two different NARX structures, according to the available data. In the first model (M1) the recruitment is forecasted from large scale oceanographic conditions in the EPO as explanatory variables. This model uses the spawners biomass data, the Southern Oscillation Index (SOI), and the Pacific Decadal Oscillation (PDO) index (Mantua et al., 1997) as inputs for the period from 1975 to 2012, a common period for all these variables. The second model (M2) consists of spawners biomass data, sea surface temperature, wind magnitude, and magnitude of the sea water currents as inputs. In this case the shorter period from 1980 to 2012 was used, since the input oceanographic variables were available only for this time span. Notice that in this model (M2) these oceanographic variables represent conditions in the area where more juvenile catch is observed within the EPO. This area was chosen based on the map of catches by fishing-set developed by the Interamerican Tropical Tuna Commission (IATTC, 2015). We chose the area of the highest YFT catch associated with floating objects, assuming that there the highest recruitment occurs, since the smallest organisms (b 60 cm) are obtained from fishing around floating objects (Dreyfus-León and Robles-Ruiz, 2008; IATTC, 2013). It extends from 1°N, 15°S and 75–95°W (Fig. 1). Furthermore, in order to produce YFT recruitment forecasting we developed a variant of model M2, which includes only the oceanographic parameters and excludes the spawners biomass. These oceanographic parameters were taken for future climate-change scenarios, described in Section 2.2, but no predictions for spawners biomass were available, hence this biotic variable was excluded in this model. Nonetheless, as will be shown in the following sections, including only the abiotic, oceanographic variables in a NARX-based model results in an adequate estimate of recruitment in the EPO.

107

2.2. Data source The Stock expected Recruitment and spawners biomass data were provided by the IATTC. These estimates come from the Stock Synthesis model 3.23b (Methot, 2011), a statistical age-structure model for stock assessment. SOI data come from the National Center for Atmospheric Research (NCAR), available on http://www.cgd.ucar.edu/cas/catalog/ climind/soi.html. PDO index data was obtained from the Join Institute for the Study of the Atmosphere and Ocean (JISAO) from its website: http://jisao.washington.edu/pdo/. Wind data were obtained from a monthly reanalysis product with a 2.5° spatial resolution, held by the Physical Sciences Division (PSD) of the National Oceanic and Atmospheric Administration (NOAA) available through its website: http:// www.esrl.noaa.gov/psd/. Currents velocity data was taken from a monthly reanalysis product with an 1/3° resolution in latitude and 1° in longitude, provided by Global Ocean Data Assimilation System (GODAS), also available in the previous website. Sea surface temperature monthly data were obtained with a 2° resolution, derived from an analysis of the Comprehensive Ocean-Atmosphere Data Set (ICOADS) provided by NOAA in its website: http://icoads.noaa.gov/products.html. Recruitment forecast was carried out by using oceanographic parameters (sea surface temperature, wind magnitude, magnitude of the sea water currents) from four different climate-change scenarios as inputs. These scenarios are numerical-modeling future simulations forced by specific emissions of greenhouse gases, referred as representative concentration pathway (RCP): RCP2.6, RCP4.5, RCP6.0, and RCP8.5 (IPCC, 2013). RCP labels then provide rough estimates of radiative forcing that will be achieved through the year 2100. Numerous General Circulation Model (GCM) outputs, available in the Coupled Model Intercomparison Project Phase 5 (CMIP5) website (http://cmip-pcmdi. llnl.gov/cmip5/), provide atmospheric and oceanic variables which are used for the assessment of the climate change impacts carried out by the Intergovernmental Panel on Climate Change (IPCC; http://www. ipcc.ch/). Herein the GCM we chose is the Geophysical Fluid Dynamics Laboratory (GFDL) Coupled Physical Model CM3, available in the GFDL website: http://data1.gfdl.noaa.gov. We decided to use this GCM (GFDL CM3) because since its previous versions (GFDL 2CM.0 and GFDL CM2.1) it has proved to reproduce the oceanic dynamics associated with the El Niño (Lin, 2007) and the PDO (Overland and Wang, 2007). 2.3. NARX configuration

Fig. 1. Area of the highest YFT catch associated with floating objects, on the map of the long-term annual mean sea surface temperature (in °C).

NARX models were configured using nnstart tool in Matlab R2012b. A neural network consists of an input layer composed of processing elements (neurons) that receive input signals at the start of the network, one hidden layer, and an output layer with one output neuron. NARX is a recurrent dynamic network, with feedback connections, where both the input layer and the hidden layer receive inputs signals (x) at time t but also at time t-n, where n is the number of lags. In the same way receive as input signals, the output (y) calculated at time t-n (Beale et al., 2014). The network configuration consists of determining the number of neurons in the hidden layer and the number of lags in the input signals to maximize modeling ability. There is no default criterion to determine the best structure, therefore we evaluated the performance of the network with different structures, after a training phase. In this training phase a set of input/target pairs are used for the back propagation algorithm training (Rumelhart et al., 1986). The best models were selected according to the criteria in Beale et al. (2014). In addition, we calculated the correlation coefficient between the estimated recruitment and the expected recruitment (estimated recruitment provided by IATTC) for the entire data set (General CC), and the correlation coefficient between the recruitment calculated for those input signals not used during the training (test data) and the recruitment expected for that dataset (Test CC).

108

L.K. Torres-Faurrieta et al. / Ecological Informatics 36 (2016) 106–113

2.4. Sensitivity analysis The model's sensitivity to changes in the input variables was evaluated, in order to characterize the degree of importance of each variable in its corresponding model. In M1, variables were increased and reduced 10% and 30%, from its original value. In M2 they were increased and reduced 5% and 10%. These percentages were selected trying to keep real conditions of the ocean. While an input variable was changed, the rest remained in their original values. To assess the importance of the variables, we calculated the mean square error (MSE) between the initial estimates of the model and the resulting estimates of each modification, as well as the sum of squared differences (SSD) and the correlation (CC) between them. To visualize the behavior of the estimates in the analysis, we present time series graphs or scatter plots as needed.

Table 1 Parameters of the sensitivity analysis calculated between the initial estimates and the estimates obtained with every percentage change in variables of M1 (SSD = sum of squared differences; MSE = mean square error; CC = correlation coefficient). Variable

% variation

SSD (1010 tons)

MSE (104 tons)

CC

SOI

+10 −10 +30 −30 +10 −10 +30 −30 +10 −10 +30 −30

3.17 3.36 26.80 28.30 2.23 2.08 19.10 14.80 3.17 2.60 31.10 17.30

1.46 1.51 4.24 4.36 1.23 1.18 3.58 3.15 1.46 1.32 4.57 3.41

0.98 0.98 0.86 0.90 0.99 0.99 0.93 0.92 0.99 0.99 0.94 0.95

PDO

Spawners biomass

3. Results 3.1. NARX configuration and sensitivity analysis The best performance in M1 and in M2 was obtained by configuring the network with 4 neurons in the hidden layer and 3 lags in the input variables. General CC and Test CC values obtained in M1 are r N 0.85 and r N 0.83, respectively. In M2, General CC and Test CC values are r N 0.82 and r N 0.89, respectively. Fig. 2 shows the time series of expected and modeled quarterly recruitment in M1. The parameters evaluated in the sensitivity analysis in M1 showed that the SOI is the most important variable for this model when the variables values were modified in 10%, since the highest SSD and MSE and the lowest CC values between recruitment initially estimated by the model and that estimated in the analysis are calculated. When the variables were modified by 30%, the spawning biomass proved to be the most important variable for the model, according to the SSD and MSE values. However, in this case the CC is higher than that obtained with the variation of SOI and PDO (Table 1). The neural network finds no clear relation between positive and negative changes in SOI. There seems to be interdependency with the other input variables, since changes in recruitment are both positive and negative with similar corresponding changes of SOI values (Figs. 3). By contrast in the case of spawning biomass, a positive change of this variable produces increases of recruitment estimates and vice versa (Fig. 4). The Model response to modifications in the PDO values is highly variable throughout the series, making it difficult to establish relationships between these variables and estimates of recruitment. In M2, sea surface temperature is the most important variable in the 5% change sensitivity runs. However, wind magnitude is the most

important when variables are modified by 10% (Table 2). The increase in temperature values increases significantly the recruitment estimates and vice versa (Fig. 5). Reducing the wind magnitude values increases the recruitment estimates, compared to the increased values (Fig. 6). The Model response to modifications in water-current magnitude and spawners biomass values is highly variable throughout the series, making it difficult to establish relationships between these variables and estimates of recruitment. 3.2. Recruitment future projections The forecasting model consists of 7 neurons in the hidden layer and 3 lags in the input variables. General CC and Test CC values are of r = 0.90 and r N 0.83, respectively. Fig. 7 shows the recruitment forecasting for the period 2014–2090, estimated from the 4 different RCP scenarios. In general, long-term estimations (throughout the study period) of recruitment are higher under RCP8.5 conditions, and lower under RCP2.6 conditions. The farther we move to the future to forecast recruitment based on this RCP scenarios, more uncertainty exists that they represent the future state of nature. For that reason a short term average recruitment prediction is calculated (period 2015–2020). The RCP2.6 is the most favorable scenario, because on average 188,465 tons of recruits are estimated. RCP 6.0 with 162,856 tons of recruits is the least favorable scenario. For the RCP4.5, 174,555 tons of recruits are estimated, and 173,525 for RCP 8.5. 4. Discussion 4.1. Influence of oceanographic parameters in recruitment

Fig. 2. Observed and modeled recruitment in M1.

ANN's provide an alternative way to model nonlinear dynamic systems such as recruitment of YFT. Variations in some input variables produced different qualitative changes in the predicted variable. We showed that YFT recruitment can be modeled from oceanographic and biological parameters. Sensitivity analysis is useful to assess the importance of variables in the model estimates. Furthermore, this analysis can provide insights to deduce effects of each variable on real recruitment. However, the information provided by the sensitivity analysis regarding these relations must be handled cautiously. SOI is the most important variable in M1, although increases in SOI did not always result in increases of Recruitment predicted values. It is possible that, since M1 considers general conditions in the eastern Pacific, the relation is somehow hidden due to averaging SOI conditions of different areas. The relationship between SOI and YFT recruitment has been previously reported by some authors (e.g., Joseph and Miller, 1989; Lehodey et al., 2003). According to Lehodey et al. (2003), a positive phase of the PDO (which is characterized by conditions of El Niño) has positive impact on the recruitment of yellowfin tuna. PDO is of minor importance for the model, so modifying their values causes

L.K. Torres-Faurrieta et al. / Ecological Informatics 36 (2016) 106–113

109

Fig. 3. Scatterplot of the sensitivity analysis for the SOI in M1. Panels (a) and (b) correspond to ±10% variations; panels (c) and (d) correspond to ±30% variations.

no significant differences in the estimates. We suppose that our model is less sensitive to signals of decadal scale, given the limited length of the input time series. Although no apparent relation between spawners and recruitment for YFT has been found, M1 model establishes some relation between those variables. It is possible that this relation is real and shaped at the planktonic YFT stage. That was a major reason to use the NARX-ANN configuration with several time lags for the input variables. In the same way, temperature is the most important variable in M2. In this model, an increase in temperature always increases the Recruitment output estimate. In contrast to M1 (in particular SOI sensitivity results), M2 model uses “local” input variables. The results are consistent with reports that YFT has a marked thermal preference in its geographical distribution (Collete and Nauen, 1983), in its spawning areas (Margulies, 2001), and the optimal growth of larvae (Wexler et al., 2011). The increase in temperature increases the recruitment estimates, which is consistent with the fact that El Niño years are considered positive for recruitment. According to Margulies et al. (2007), temperature has important effects on larval stages of YFT; low temperatures prolong the larval stage developing period making larvae more susceptible to depredation. The optimal temperature range for fast growth and high survival of larvae of yellowfin tuna is 26 °C to 31 °C (Wexler et al., 2011). According to Holey and Maunder (2007) and Langley et al. (2009), there is a weak spawning stock-recruitment relationship. Herein this could be seen only if the biomass drops to very low levels, which implies a significant change in their values. In contrast, M1 model found a relatively strong stock-recruitment relation not only for decreased adult abundance values, and in contrast to SOI where an interaction with other input variables is clearly seen in output values, variations in

spawning biomass produce direct changes in recruitment estimates. The effect of this variable is not altered drastically by the other input variables. The wind magnitude has an important impact on recruitment estimates on M2. This notion is reasonable, since the wind is responsible for the onset of coastal upwelling and generation of turbulence, which influence in the larvae feeding, and consequently the survival and recruitment. The largest recruitment area included in this model corresponds to a part of the Peruvian upwelling costal system, a strong wind area that reaches monthly-mean magnitudes of 7.8 m s−1 in the study period. Langley et al. (2009) also associated high recruitment of YFT tuna to an area of strong winds in the western and central Pacific Ocean. According to Cury and Roy (1989), the annual rate of recruitment of pelagic fish increases with the intensification of upwelling, until the wind reaches speeds about 5–6 m s−1. Haury et al. (1990) and Kimura et al. (2004) also suggest that there is an optimum wind range that induces moderate upwelling, which is strong enough to promote the primary production and to raise the encounter rate of larvae with prey, but weak enough to avoid dispersion of feeding larvae patches. It is possible that the reduced values in the sensitivity analysis favored the recruitment estimates, because values in this region are higher than that interval suggested by Cury and Roy (1989). Modification of the currents magnitude causes no significant differences in the estimation of recruitment, therefore it is not possible to establish a clear link between this variable and the recruitment. ANN's models have been considered as black boxes (Kriesel, 2007; Lek-Ang et al., 1999; Zhang et al., 1998) implying that there is no explicit way to clarify the relationship between inputs and outputs in these models, which causes difficulty in interpreting the results. However,

110

L.K. Torres-Faurrieta et al. / Ecological Informatics 36 (2016) 106–113

Fig. 4. Scatterplot of the sensitivity analysis for the spawners biomass in M1. Panels (a) and (b) correspond to ±10% variatrions; panels (c) and (d) correspond to ±30% variations.

the sensitivity analysis proposed in this study is a good approach to assess the degree of real influence of an input variable on the calculated output, since the most important variables to the yellowfin tuna recruitment highlighted in the results are those variables that have been categorized as such in the literature, by modeling and laboratory experiments.

Table 2 Parameters of the sensitivity analysis calculated between the initial estimates and the estimates obtained with every percentage change in variables of M2 (SSD = sum of squares differences; MSE = mean square error; CC = correlation coefficient). Variable

% variation

SSD (1010 tons)

MSD (104 tons)

CC

Temperature

+5 −5 +10 −10 +5 −5 +10 −10 +5 −5 +10 −10 +5 −5 +10 −10

35.7 37.7 132.0 142.0 12.8 25.2 32.8 109.0 0.5 0.6 2.10 2.30 0.3 0.4 1.3 1.6

5.32 5.47 10.2 10.6 3.18 4.47 5.10 9.31 0.65 0.67 1.29 1.35 0.52 0.54 10.2 11.1

0.95 0.94 0.82 0.78 0.92 0.89 0.79 0.67 0.99 0.99 0.98 0.98 0.99 0.99 0.99 0.99

Wind magnitude

Spawners biomass

Currents magnitude

4.2. Recruitment forecasting Although the M1 model uses spawners biomass to shape recruitment estimates, the forecasting model was implemented only based on predicted oceanic conditions because it is impossible in the long run to know how the fishery will impact the stock, and because it is considered that at the levels of biomass that the YFT has experienced with fishery interaction, there is no relation between spawners biomass and the recruitment they produce each year (IATTC, 2015). The different RCP scenarios generate emission projections of various atmospheric constituents, such as greenhouse gases and aerosols (Taylor et al., 2012). Although scenarios are under certain assumptions, they are the support of the climate change research of the IPCC. The model designed for recruitment forecasting of YFT has the ability to give a potential recruitment response to the different RCP's. The long-term future projection estimates shows that recruitment is higher for the most radiative forcing scenario (RCP8.5), and lower for the least radiative forcing scenario (RCP2.6). It is possible that this response is due to the temperature associated with each scenario. The average temperature is highest under RCP 8.5 conditions (given more radiative forcing), which falls around ~27 °C by the end of the study period, while the average temperature is lowest under RCP2.6 conditions and falls around ~24 °C by the end of period. These results support the idea that high temperatures are favorable for the yellowfin tuna recruitment, as reported in the sensitivity analysis of M2. The short-term estimates differ from previous results, since the recruitment is higher under RCP2.6 conditions in which the temperature is slightly higher than in other scenarios. RCP 2.6 is the

L.K. Torres-Faurrieta et al. / Ecological Informatics 36 (2016) 106–113

111

Fig. 5. Scatterplot of the sensitivity analysis for the sea surface temperature in M2. Panels (a) and (b) correspond to ±5% variations, the panels (c) and (d) correspond to ±10% variations.

so called “peak-and-decay” scenario, in which the radiative forcing reaches a maximum (hence also the temperature) near the year 2050, before decreasing to a level of 2.6 W m−2 in the year 2100. Despite the assumptions on which RCP's projections are based on, it is likely that the behavior of the oceanographic variables, particularly in a short-term basis, will be in the interval of values calculated under different scenarios, because they are based on historical emissions and scenarios proposed in the literature (Thomson et al., 2011), hence the ANN's are a good tool to make potential forecasting. According to IPCC reports (IPCC, 2013), the EPO shows a warming trend which in compliance with the biological response of YFT, and based on the results of this study, these conditions could favor the recruitment. However, for more accurate recruitment forecasting it is essential to consider the effect of such a warming on the ecosystem as well as the status and management of the resource. This study shows that NARX are a viable tool for ecological modeling. The forecasting model is able to make long time projections; however, the uncertainty grows as it progresses over time, coupled with the uncertainty of the recruitment initial estimates that feeds the input layer of the model. Therefore, we consider that only a short period would result in a reasonable forecast. Based on the different scenarios, the projections show that the recruitment will be in the range of 162,856 to 188,465 tons in annual average in the period 2015–2020. These estimates in any of those scenarios represent an increase in the YFT recruitment for the following years, relative to that estimated by the IATTC in the period 2009–2013, which was of 142,008 tons. These results represent valuable information that can be useful for management consideration in relation to

conservation measures for YFT. There is also a need to understand the response of the ecosystem to climate change, so according to this model, YFT would benefit by expected oceanographic conditions, although other components of the ecosystem would suffer changes that might have an impact on YFT with unexpected results.

5. Conclusions The recruitment of yellowfin tuna in the eastern Pacific Ocean is modeled using nonlinear autoregressive network models with exogenous inputs (NARX), with either teleconnection indices (SOI and PDO) or regional oceanographic fields (temperature, wind, and currents) as explanatory variables in two different approaches (Models 1 and 2), respectively. Both NARX-model approaches produce similar results, reproducing most of the variability observed in the “real” yellowfin recruitment, with correlation coefficients N 0.80. The sensitivity analysis shows that the SOI and the sea surface temperature are indeed the leading variables in the modeled recruitment, most probably associated to the predisposition of the yellowfin-tuna larvae to thrive in a thermal range of 26 °C to 31 °C. The future projections of the recruitment, using climate-change scenarios, predict generally more favorable conditions for the recruitment, given the higher temperatures projected in the Pacific. Nonetheless, this result should be interpreted with caution, since the NARX model considers no other factors of paramount importance like the environmental effects on the tuna's preys or impacts from the fishery itself; this topic deserves further research and should be addressed in future analyses.

112

L.K. Torres-Faurrieta et al. / Ecological Informatics 36 (2016) 106–113

Fig. 6. Scatterplot of the sensitivity analysis for wind magnitude in M2. Panels (a) and (b) correspond to ±5% variations; panels (c) and (d) correspond to ±10% variations.

Acknowledgements We thank the anonymous reviewers for their comments and suggestions to and earlier version of this manuscript. This research was supported by Mexico's CONACYT through a graduate scholarship to KTF. DR has been funded by CICESE's budget through the internal project 625118. Comments and suggestions during the first stages of this

Fig. 7. Recruitment projections in scenarios with different RCP's.

research by Oscar Sosa Nishizaki and Enric Pallás Sanz enriched this work and are thankfully acknowledged. References Beale, M.H., Hagan, M.T., Demuth, H.B., 2014. Neural Network ToolboxTM User's Guide. Chen, I., Ware, D.M., 1999. A neural network model for forecasting fish stock recruitment. Can. J. Fish. Aquat. Sci. 56 (12), 2385–2396. Collete, B., Nauen, C., 1983. FAO species catalogue. Scombrids of the World. An Annotated and Illustrated Catalogue of Tunas, Mackerels, Bonitos and Related Known to Date. FAO Fisheries Synopsis vol. 2 (125 p). Cury, P., Roy, C., 1989. Optiman enviromentl window and pelagic fish recruitment success in upwelling areas. Can. J. Fish. Aquat. Sci. 46 (4), 670–680. Dreyfus-León, M., Robles-Ruiz, H., 2008. Tallas (Longitud furcal) del Atún Aleta Amarilla, Thunnus albacares, capturado por la flota atunera mexicana durante 2007. El Vigía. PNAAPD, pp. 5–7 (Enero-Junio, año 13 num. 33). Dreyfus-León, M., Schweigert, J., 2008. Recruitment prediction for Pacific herring (Clupea pallasi) on the west coast of Vancouver Island, Canada. Ecol. Inform. 3 (2), 202–206. Haury, L.R., Yamazaki, H., Itsweire, E.C., 1990. Effects of turbulent shear flow on zooplankton distribution. Deep Sea Res. Part A Oceanogr. Res. Pap. 37 (3), 447–461. Holey, S.D., Maunder, M.N., 2007. Status of yellowfin tuna in the Estern Pacific Ocean in 2005 and outlook for 2006. Inter-American. Tropical Tuna Commission., Stock Assessment Report. 7, pp. 3–115. IATTC, 2013. Tunas and Billfishes in the Eastern Pacific Ocean in 2012. Fisheries Status Report 11. (La Jolla Cal.). p. 171. IATTC, 2015. Tunas, Billfishes and Other Pelagic Species in the Eastern Pacific Ocean in 2014. (Doc IATTC-8904a presented in IATTC annual meeting, Guayaqyil, Ecuador, june 2015. http://www.iattc.org/Meetings/Meetings2015/June/PDFs/IATTC-89-04aTunas-billfishes-and-other-pelagic-species-in-the-EPO-2014.pdf). IPCC, 2013. Climate Change 2013. In: Stocker, T.F., Qin, D., Plattner, G.K., M., T., K., A.S., J., B., A., N., Y., X., V., B., M., y.M.P. (Eds.), The Physical Science Basis Contribution of Working Group I to the Fith Assesment Report of the Intergovernmental Panel on Climate Change. Cambridge University Press, Cambridge, United Kingdom and New York, NY, USA, p. 135. Jarre-Teichmann, A., Brey, T., H., y.H., 1995. Exploring the use of neuronal network for forecast in the Peruvian upwelling. NAGA the ICLARM Quarterly. 18(4), pp. 38–40.

L.K. Torres-Faurrieta et al. / Ecological Informatics 36 (2016) 106–113 Joseph, J., Miller, F.R., 1989. El Niño and the surfce fishery for tunas in the eastern Pacific. Bull. Jpn. Soc. Fish. Oceanogr. 57, 77–80. Kriesel, D., 2007. A Brief Introduction to Neural Networks. (Retrived from http://www. dkriesel.com/en/science/neural_networks). Kimura, S., Nakata, H., Margulies, D., Suter, J.M., Hunt, S.L., 2004. Effect of oceanic turbulence on the survival of yellowfin tuna larvae. Nippon Suisan Gakkaishi 70 (2), 175–178. Langley, A., Briand, K., Kirby, D.S., Murtugudde, R., 2009. Influence of oceanographic variability on recruitment of yellowfin tuna (Thunnus albacares) in the western and Central Pacific Ocean. Can. J. Fish. Aquat. Sci. 66 (9), 1462–1477. Lehodey, P., Chai, F., Hampton, J., 2003. Modelling climate-related variability of tuna populations from a coupled ocean-biogeochemical-populations dynamics model. Fish. Oceanogr. 12 (4–5), 483–494. Lek-Ang, S., Deharveng, L., Lek, S., 1999. Predictive models of collembolan diversity and abundance in a riparian habitat. Ecol. Model. 120 (12), 247–260. Lek, S., Guégan, J.F., 1999. Artificial neural networks as a tool in ecological modelling, an introduction. Ecol. Model. 120 (2), 65–73. Lin, J.L., 2007. Interdecadal variability of ENSO in 21 IPCC AR4 coupled GCMs. Geophys. Res. Lett. 34 (12). Mantua, N.J., Hare, S.R., Zhang, Y., Wallace, J.M., Francis, R.C., 1997. A Pacific interdecadal climate oscillation with impacts on salmon production. Bull. Am. Meteorol. Soc. 78 (6), 1069–1079. Margulies, D., 2001. Early life history studies of yellowfin tuna, (Thunnus albacares). InterAm. Trop. Tuna Comm., Bull. 22, 9–20. Margulies, D., Sutter, J.M., Hunt, S.L., Olson, R.J., Sholey, V.P., Wexler, J.B., Nakazawa, A., 2007. Spawing and early development of captive yellowfin tuna, (Thunnus albacares). Inter-Am. Trop. Tuna Comm., Bull. 105 (2), 249–265.

113

Methot, R.D., 2011. User's Manual for Stock Synthesis: Model Versión 3.23b. (November 7, 2011). NOAA Fisheries Service, Seattle Washington. Overland, J.E., Wang, M., 2007. Future climate of the North Pacific Ocean. Eos Trans. AGU 88 (16), 178–182. Rumelhart, D.E., Hinton, G.E., Williams, R.J., 1986. Learning Internal Representations by Error Propagation. In: Rumelhart, D.E., McClelland, J.L. (Eds.), Parallel Data Processing vol. 1. The M.I.T. Press, Cambridge, MA, pp. 318–362. Shaefer, M.K., 1998. Reproductive biology of yellowfin tuna (Thunnus alabacares) en the eastern Pacific Ocean. Inter-Am. Trop. Tuna Comm., Bull. 21 (5), 272. Taylor, K.E., Stouffer, R.J., Meehl, G.A., 2012. An overview of CMIP5 and the experiment design. Bull. Am. Meteorol. Soc. 93 (4), 485–498. Thomson, A.M., Calvin, K.V., Smith, S.J., Kyle, G.P., Volke, A., Patel, P., Edmonds, J.A., 2011. RCP4. 5: a pathway for stabilization of radiative forcing by 2100. Clim. Chang. 109 (1– 2), 77–94. Wexler, J.B., Margulies, D., Scholey, V.P., 2011. Temperature and dissolved oxygen requirements for survival of yellowfin tuna, Thunnus albacares, larvae. J. Exp. Mar. Biol. Ecol. 404 (1), 63–72. Yáñez, E., Plaza, F., Gutiérrez-Estrada, J.C., Rodríguez, N., Barbieri, M.A., Pulido-Calvo, I., Bórquez, C., 2010. Anchovy (Engraulis ringens) and sardine (Sardinops sagax) abundance forecast off northern Chile: a multivariate ecosystemic neural network approach. Prog. Oceanogr. 87 (1), 242–250. Zhang, G., Patuwo, B.E., Hu, M.Y., 1998. Forecasting with artificial neural networks: the state of the art. Int. J. Forecast. 14 (1), 35–62.