| Chilean jack mackerel( Trachurus murphyi) is an important marine economic species, and a main fishing species of large pelagic trawlers. Because its fishery resources distribution across all the South Pacific and easy to be affected by the marine environment, it’s hard to master the central fishing grounds and resource abundance.Using some advanced fishery forecasting method to accurately predict the central fishing grounds and resource abundance will be helpful to improve the fishery production efficiency and provide some reference for scientific management of fisheries management organizations of resources and sustainable utilization.According to the statistical data of the production of 15 fishing vessels fishing in the Southeast Pacific fishing ground from 2003 to 2013, and 7 remote sensing marine environmental factors include of sea surface temperature, sea surface temperature gradient, sea surface height, chlorophyll a concentration, Southern Oscillation Index,Nino 3.4 index and its anomalous value. Through the analysis with year、month、vessel、longitude、latitude and latitude and longitude interaction factor from production statistical data and 7 marine environment factors by GAM, the non significant factor of the sea surface temperature gradient factor was eliminated. We use the other factors to building standardized CPUE model which make CPUE as resource abundance index.The time scale of the standardized CPUE data sample is the month, and the spatial scale is 1°×1°. Then, 9 time-space and environment factors that the longitude, latitude,Southern Oscillation dynamic index of sea surface height, chlorophyll a concentration,sea surface temperature, Nino3.4 index and its anomalous value will be used as input layer factors of prediction system based on BP neural network, RBF neural network,GABP neural network. The standardized CPUE will be treated as output layer factor.The number of hidden layer is 1. We could use the linear fitting goodness R2 index,residual sum of squares and other statistics parameters to study and obtained the optimal neural network model in three kinds of neural network.(1) Standardized CPUEThe fishing efficiency of 15 fishing vessels was significantly different. Fishing vessel factor effect explanation was 7.91%; The CPUE present S type fluctuations overall through 2003 to 2013.CPUE rise in 2003~2006, and slowly decline in2007~2012, but rise again in 2013, the year factor effect explained 7.79% of the model.The abundance of resources has obvious seasonal fluctuation. It gradually rises from Jan. to Apr., maintain stable high yield in May., Mar., Jul., and then decreased until the Jan. of next year from Sep., the month factor effect explanation is 13.7%. The cumulative effect of the three accounted for 76% of the total explanatory power of the model, which had a significant impact on the resource abundance of Chilean jackmackerel. In the spatial factors part, the main fishing ground located in 45S~32S and 80W-100 W, center fishing ground presents trend from west to east from the north to the south, the spatial factors effect explanation is 3.3%. In the marine environment factors part, the most suitable habitat sea surface temperature was11~17℃;the most suitable habitat chlorophyll a concentration was 0.1~0.2mg/m3; When the Nino3.4index is26.5~28 or its anomalous value is-0.5~0.5, the SOI index is-20、-5、5、12、30, the sea surface height is-10cm、10cm、40cm,the fishing ground will get higher abundance of resources.Through the AIC test by adding sea surface gradient factor to GAM, the explanation was declined, and the significance was weak. Therefore, we did not account this factor as the model factor when constructing the standardized CPUE model based on GAM. The marine environment factors effect explanation is 3.6%.The final standardized CPUE model based on GAM is as following:ln(cpue+0.9)~factor(vessel)+factor(year)+factor(mon)+s(lat)+s(lon)The totally model explanation was 38.7%, and the standardized CPUE could reflect thetrend of the characteristics change in the nominal CPUE.(2) BP neural network fishery forecasting system analysisFor the BP neural network fishery forecasting system of the Southeast Pacific Chilean jackmackerel, with the increase of the numbers of single hidden nodes from1 to 16, firstly, the accuracy of the forecasting increases and then decreases slightly and gradually stable and slight fluctuations. Less hidden layer nodes will lead to neural network model structure too simple, too little iteration, too small convergence;however, too large implicit layer node number will enhance the ability of model mapping, greatly prolonged learning time, and even over fitting. Comprehensive the residual squares and fitting optimization coefficients,we found that the single hidden layer neural network nodes are more suitable for the 6.At this time, goodness of fit coefficient is 0.5202, residual is 0.3167, standard deviation is 0.038. The forecasting model prediction results basically reflected the observation of the standardization of CPUE site changes and meet forecast demand except for the local area appeared to fall into a small value.(3) RBF neural network fishery forecasting system analysisFor the RBF neural network fishery forecasting system of the Southeast Pacific Chilean jackmackerel, with the increase of the numbers of single hidden nodes from4 to 20, the accuracy of the forecasting increases and then gradually stable. Different nodes will fit different best spread coefficient. For the forecasting model result will be stable with the same model parameters, it will be the most important concerned influence factor when we choose a suitable algorithm to calculate appropriate sample set center. By the comparison analysis of the fishery forecasting model parameters, it was concluded that when the hidden nodes was 14, spread coefficient 0.6, the accuracy of forecasting and residual sum of squares would be 29.07% and 89. The forecasting model prediction results could only.The forecasting model prediction results roughly reflected the observation of the standardization of CPUE site changes. The overall forecast value was higher than the observed standard CPUE value.(4) GABP neural network fishery forecasting system analysisFor the GABP neural network fishery forecasting system of the Southeast Pacific Chilean jackmackerel, with the increase of the sizes of population, the accuracy of theforecasting increases and then gradually stable. The best population size was 250.However, the effect of the mutation rate and crossover rate on the performance of the prediction model is great. Too much mutation rate and crossover rate could lead to instability of prediction, but too small can lead to the model couldn’t reflect the characteristics of the complexity of the fishing ground well, the model would be too convergent. Through the comparison, we found that when the mutation rate was 0.09,the crossover rate 0.75, population size 250, hidden nodes 6, the accuracy of forecasting and residual sum of squares would be 69.52% and 13.5864,and it could to be the best forecasting model. The forecasting model prediction results reflected the observation of the standardization of CPUE site changes and source abundance of the central fishing ground.(5) Three neural network fishery forecasting system comparison analysisThrough the performance comparison of the three models, the standard BP and RBF neural network showed general in fishery forecast of Chilean jackmackerel. The BP neural network performed a litter better than RBF neural network. However,the GABP neural network performed much better than BP neural network with the forecasting accuracy and stability. The BP neural network forecasting system appeared local into the minimum value in the site 25-75 and the RBF neural network forecasting system appeared too high forecast value due to the selection of the center not reasonable. For the Optimization selection of parameters in global search based on genetic algorithm of GABP neural network, the whole forecasting model could be more stable. And there are only a bit deviation occurred when using the CPUE overlay graph of 2012. It indicate that the application of GABP neural network in the Chilean jackmackerel fishery forecasting was feasible. And it would provide a new reference method and research angle for the future study of the neural network model in the fishery forecasting. |
PDF Full Text Request