基于BP神经网络的西北太平洋柔鱼资源丰度预测

柔鱼是具有巨大开发潜力的重要经济头足类种类，广泛分布在太平洋海域。柔鱼是短生命周期种类，其生活史过程与栖息地的海洋环境条件有重要关联，海洋环境因子的时空分布与变化显著影响着柔鱼资源的分布范围和资源密度。本文基于监督式学习算法的BP神经网络模型，综合多源卫星遥感观测获取得到的海表温度（Sea surface temperature；SST）、叶绿素a浓度（Chlorophyll-a concentration；Chl-a）、海表面高度距平（Sea surface height anomaly；SSHA）、海水质量变化和地转流等海洋环境因子，对西北太平洋柔鱼资源丰度的时空分布进行了模拟和预测。以上海海洋大学中国远洋渔业数据中心2004~2017年的西北太平洋海域的柔鱼历史渔业捕捞数据为参考值，对基于多源卫星遥感观测的多海洋环境因子的柔鱼资源丰度的模拟和预测结果进行了精度评定。结果表明，与仅采用SST、Chl-a和SSHA等进行柔鱼资源丰度时空分布预测的传统方案相比，进一步加入海水质量变化和地转流后，可有效提高利用BP神经网络对西北太平洋柔鱼资源丰度进行模拟和预测的精度：改进方法模拟的标准差（Standard deviation; STD）和均方根误差（Root mean square error; RMSE）均提高了22%；且预测的STD提高了31%，RMSE减少了26%。

Prediction of the CPUE of neon flying squid in the northwest Pacific Ocean based on back propagation neural network

The neon flying squid, Ommastrephes bartramii, is one of the important cephalopod with great potential for economic development, which is widely distributed over Pacific Oceans. O. bartramii has a short lifespan, and its life history stages is affected by the ambient oceanographic regimes and the epipelagic environment, and its spatial and temporal distributions is highly related to the variability in various oceanographic variables. By incorporating the sea surface temperature (SST), chlorophyll-a concentration (Chl-a), sea surface height anomaly (SSHA), ocean mass and geostrophic current from multi-source remote sensing observations, the supervised learning algorithm-based back propagation (BP) neural network model is implemented in this paper to model and predict the temporal-spatial distributions of the catch per unit effort (CPUE) of O. bartramii in the Northwest Pacific Ocean. The multi-source remote sensing data are used to build the BP neural network model, and the accuracy of the model-simulated and -predicted O. bartramii CPUE is then evaluated with the historical fishery data during 2004-2017 in the Northwest Pacific Ocean from the Chinese Squid-Jigging Technology Group of Shanghai Ocean University. Comparing with the traditional scheme of predicting the spatial-temporal distributions of O. bartramii CPUE with SST, Chl-a and SSHA only, the accuracy of model-simulated and -predicted CPUE can be further improved after incorporating the ocean mass and geostrophic current into the BP neural network model. Specifically, the standard deviation (STD) and root mean square error (RMSE) of model-simulated O. bartramii CPUE are both decreased by 22%, as well as those of model-predicted O. bartramii CPUE are reduced by 31% and 26%, respectively.