基于不同BP神经网络的西南大西洋阿根廷滑柔鱼渔场预报模型比较

Comparison of Different Forecasting Model for Fishing Ground of Illex argentinus Based on Artificial Neural Networks in the Southwest Atlantic Ocean

根据2003-2011年渔汛期间我国鱿钓船在西南大西洋海域的生产统计数据,结合海洋遥感获得的海表温度(SST)和海面高度(SSH)等数据,以单位捕捞努力量渔获量(CPUE)和作业次数作为中心渔场指标,以月份、经度、纬度、SST和SSH为输入因子,利用BP神经网络方法构建西南大西洋阿根廷滑柔鱼中心渔场预报模型。比较14种不同结构的BP神经网络模型,以CPUE作为中心渔场预报指标的BP模型均较佳,其拟合残差范围为0.004 0~0.005 5,平均值为0.004 7;而以作业次数作为中心渔场预报指标的BP模型,其拟合残差范围为0.009 3~0.011 6,平均值为0.010 4。输入因子为月份、经度、纬度、SST和SSH,输出因子为初值化后的CPUE,网络结构为5-4-1时的BP神经网络模型为最佳,其拟合残差为0.004 025,该模型可用于阿根廷滑柔鱼中心渔场的预报。BP神经网络方法可为准确渔场预报提供新途径。

Illes argentinus is an important economic cephalopod in the world. Its distribution of fishing ground is closely related to the marine environmental factors, and the accurate forecasting of fishing ground can provide better scientific guidance for fishing operation. This paper built the forecasting models by using the methods of BP neural networks with the catch data from the Chinese squid jigging fleet, and the remote sensing data including sea surface temperature (SST) and sea surface height (SSH)from November to August of the next year during the years of 2003 to 2011. The BP neural networks model was applied to predict the distribution of fishing grounds of Illex argentinus based on the standardization of CPUE and fishing effort. In order to compare the BP models, fishing month, fishing position (latitude and longitude), SST and SSH were randomly used as inputting factors. The total of 14 models with different hidden layers had been tested and compared. The optimum model was selected by comparing the values of simulation residual of each model. The result showed that the BP neural networks model with standardization of CPUE was better, in which the simulation residual ranged from 0.004 7 to 0.005 7, and its mean value was 0.005 2. While the BP neural networks model with standardization of fishing effort was used, the simulation residual was from 0.010 2 to 0.010 5 and its mean value was 0.010 4. It was found that the model with 5-4-1 networks structure with the lowests imulation residual of 0.004 784 could be used for better predicting fishing ground of Illex argentinus in the southwest Atlantic, in which the standardization of CPUE was considered as the output factor, and month, latitude, longitude, SST and SSH were used as the inputting factors. It was found that the BP neural networks model could provide a new way for the accurate forecasting of fishing ground.