GLM模型和回归树模型在CPUE标准化中的比较分析

在渔业资源评估中,CPUE(catch per unit effort)标准化是基础性工作。一般线性模型(generalized linear model,GLM)已成为CPUE标准化的基本方法,但GLM模型在误差结构、自变量的选择、缺失数据、复杂交互效应及异常值处理等方面仍然缺乏灵活性。本文基于模拟数据及我国东、黄海鲐鱼(Scomber japonicus)灯光围网渔业数据,比较和分析了基于GLM模型与回归树模型在CPUE标准化中的效果。研究表明:当渔业数据不存在非线性关系与异常值时,GLM模型与回归树模型均能较好地对CPUE进行标准化,但由于回归树模型具有阶跃函数特征,因而GLM模型更具优势;在非线性关系及异常值存在的条件下,回归树模型对CPUE的标准化具有相对较小的估计误差,模型更简约、有效。由于回归树模型能可视化显示自变量与应变量间的复杂关系,因此,更有利于探索和分析渔业数据。

Comparisons of regression tree and GLM performance in CPUE standardization

CPUE (catch per unit effort) standardization is an essential task in fisheries stock assessment and GLM (generalized linear model) which has been used as a standardized method in the CPUE standardization. Before using GLM, the error structure, independent variables, and interaction between variables in the model had to be assigned and it would cause a great error if the assumption was wrong. Moreover, GLM could not be used to handle missing values automatically and to detect and extract complex interactions from the CPUE data. Outliers also had a great impact on the results estimated by using GLM. In contrast to GLM, regression trees may do a great job to deal with the above situations. In this paper, based on simulation data and chub mackerel (Scomber japonicus) catch and effort data from Chinese lighting-purse seine fishery in the East China Sea and Yellow Sea, we compared the performance of the regression tree and GLM in the CPUE standardization and the results showed that both models could do a good job if there were no outliers in the data and nonlinear relationships between nominal CPUE and abundance. Because the regression tree was characterized by a step function, the GLM was better in standardizing CPUE in this situation. However, if there were outliers and nonlinear relationships, the regression tree would harvest less root mean square errors and explain more deviations with fewer variables than GLM. The regression tree also could detect the complex relationships between independent variables and response variables by visualization which was ideally suited to explore and analyze the catch and effort data from fisheries.