基于深度卷积嵌入式聚类（DCEC）的海洋环境特征提取对渔情预报模型的改进研究以西南印度洋大眼金枪鱼为例

为提高大眼金枪鱼(Thunnus obesus)延绳钓渔情预报模型的预测能力，本研究提出了一种基于深度卷积嵌入式聚类(DCEC)的海洋环境时空特征提取方法，结合广义可加模型(GAM)对西南印度洋大眼金枪鱼延绳钓渔场进行预报。采用2018年1−12月0.041 6°×0.041 6°的MODIS-Aqua和MODIS-Terra海表面温度三级反演图像数据(以日为单位)构建DCEC模型，基于Davies-Bouldi 指数(DBI)确定最佳聚类数，在此基础上提取各月海表温度(SST)的类别特征值$ {F}_{M} $；采用美国国家海洋和大气管理局网站2018年1−12月1°×1°的Chl a浓度月平均值作为辅助环境特征因子；采用印度洋金枪鱼委员会2018年1−12月1°×1°的大眼金枪鱼延绳钓渔业数据(以月为单位)，计算单位捕捞努力量渔获量(CPUE)；将SST月类别特征值$ {F}_{M} $、Chl a浓度月平均值与CPUE数据进行时空匹配，构建改进GAM；采用SST月平均值、Chl a浓度月平均值与CPUE数据构建基础GAM；采用联合假设检验($ F $检验)验证模型解释变量对响应变量的影响；采用赤池信息准则(AIC)、均方误差(MSE)、绘制实测值和预测值的散点图并计算相关系数r，分析改进GAM相比于基础GAM的提升效果。实验结果表明：(1)基于DCEC模型提取的$ {F}_{M} $能够较好地反映西南印度洋海表温度的时空动态特征与规律，并与西南印度洋的气候条件、季风状况和水文特征等相互耦合；(2) $ {F}_{M} $相比SST平均值的因子解释率更高，对大眼金枪鱼CPUE影响更为显著，高渔获率集中在暖冷流交汇区域；(3)改进GAM相比基础GAM的AIC值降低了9.17%，MSE降低了26.7%，散点图显示改进GAM预测的CPUE对数值与实测CPUE对数值的相关性较显著，r为0.60。本研究证明了DCEC模型在海洋环境特征提取方面的有效性，可为后序大眼金枪鱼延绳钓渔情预报模型的改进研究提供参考。

Improvement of marine environment feature extraction based on deep convolution embedded clustering (DCEC) for fishery forecast modelA case study of bigeye tuna (Thunnus obesus) in the Southwest Indian Ocean

In order to improve the forecasting ability of the fishery forecast model for the longline bigeye tuna (Thunnus obesus), we proposed a marine environment feature extraction method based on deep convolutional embedded clustering (DCEC), combined with generalized additive model (GAM) for forecasting the longline bigeye tuna fishing grounds in the Southwest Indian Ocean. We used the MODIS-Aqua and MODIS-Terra sea surface temperature (SST) three-level inversion image data (in days) from January to December in 2018 at 0.041 6°×0.041 6° to construct a DCEC model, determined the optimal number of clusters based on the Davies-Bouldi index (DBI), and extracted the category feature value (FM) of each month’s sea surface temperature (SST); we used monthly 1°×1° bigeye tuna longline fishery data from January to December in 2018 generated from the Indian Ocean Tuna Commission (IOTC), and calculated the catch per unit effort (CPUE); we matched the monthly category feature value FM and the monthly average value of Chl a concentration with the CPUE data to construct an improved GAM; we matched the monthly average SST, the monthly average Chl a concentration and CPUE data to build a basic GAM; we used the joint hypothesis test (F test) to verify the influence of model explanatory variables; we used akaike information criterion (AIC), mean square error (MSE), and draw the frequency distribution diagrams and box diagrams of measured and predicted values, etc., to analysis the improvement effect of the improved GAM compared to the basic GAM. The results showed that: (1) the category feature value (FM) extracted based on the DCEC model could better reflect the temporal and spatial dynamic characteristics of SST in the Southwest Indian Ocean, and was related with the climatic conditions, monsoon conditions, and hydrological characteristics in the Southwest Indian Ocean; (2) the factor interpretation of FM was higher than that of the monthly average SST in GAM, which means FM had more significant impact on the CPUE of bigeye tuna. The high catch rate was concentrated in the areas where the FM category was 2, 10, 24 with intersections between the warm and cold currents; (3) the AIC of the improved GAM was reduced by 9.17% than that of the basic GAM and MSE of the improved GAM was reduced by 26.7% than that of the basic GAM; the frequency distribution of the CPUE logarithmic value predicted by the improved GAM was closer to the normal distribution, and the high frequency distribution interval was closer to that of the measured value; the scatter plot showed that the CPUE predicted by the improved GAM had a significant correlation with the measured CPUE, with r equaled to 0.60. This study proves the effectiveness of the DCEC model in extracting marine environmental features, and can provide a reference for the further study on the bigeye tuna fishery forecast.