基于BP神经网络的中西太平洋鲣鱼渔场预报模型构建与比较

摘要

根据1998-2013年中西太平洋鲣鱼围网生产统计数据以及海洋环境数据,采用BP人工神经网络模型,分别以初值化后的单位捕捞努力量渔获量(CPUE,Catch per unit of effort)和捕捞努力量(Fishing Effort)作为中心渔场的表征因子,并作为BP模型的输出因子,以时间因子、空间因子、海洋环境因子(包括海表温度SST、海面高度SSH、Nino3.4区海表指标及叶绿素浓度Chl-a)等作为输入因子,构建22个BP神经网络模型,以最小拟合残差作为判断标准,比较渔场预报模型优劣.实验结果,以捕捞努力量为输出因子的模型的最小拟合残差均小于以CPUE为输出因子的模型,表明捕捞努力量更适合作为表征中心渔场的因子;同时,拟合残差的平均值随着输入因子的增加而减少,表明本研究所选的时间、空间、海洋环境因子等对鲣鱼中心渔场预报均极为重要.其中,以月份、经度、纬度、SST、SSH、Nino3.4a、Chl-a 为输入因子,以初值化后的捕捞努力量为输出因子,结构为7-5-1的 BP 神经网络模型预报精度为最高,影响因子的重要性从高到低依次是经度、Chl-a、SST、纬度、NINO3.4a、SSH、月份.%Central-western Pacific Ocean is an important fishing area for skipjack tuna(Katsuwonus pelamis)purse seine fisheries.Accurately forecasting the central fishing ground can help fishery fleets improve their fishing efficiency. According to the production statistical data of skipjack in the central-western Pacific Ocean during 1998 – 2013 and related oceanographic condition, 22 BP neural networks are constructed to compare the forecasting effects between Catch per unit of effort (CPUE) and fishing effort as output factors, respectively. The results showed that the value of minimum residuals in the forecasting models based on fishing effort were all lower than those based on CPUE, which represent that the fishing effort is better considered as the indicator of fishing ground. It is also found that the mean value of fitting residuals decreased as the number of input factors increased, which also show that the selected environmental variables in this study(month, sea surface temperature SST, sea surface height SSH, Nino3.4 index, Chl-a)are the important indicators affecting the distribution of skipjack. The BP neural network with a structure of 7-5-1 was the most suitable for forecasting fishing ground with a highest accuracy and lowest residuals. The significance of longitude, Chl-a, SST, latitude, Nino3.4 index, SSH and month vary from high to low.