ニューラルネットワークによる水産資源解析 : CPUE予測と要因分析の試み Fish Population Analysis by Neural Networks : Attempts for CPUE Prediction and Factorial Experiment
In the field of fish population dynamics, CPUE (catch per unit effort), defined as catch in number (or in weight) per fishing effort, is an important concept which is proportional to stock size. Because there are many areas without any operation in the fishery for southern bluefin tuna, the handling of such parts have a greatly effect on the estimated year trend of CPUE. Therefore, in this paper, we conducted the prediction of CPUE values in these missing cells using neural networks based on the spatial-temporal information. As a result of the reliability check by n-fold cross-validation, it shows rather high values of Pearson's correlation coefficient or low values of absolute error between observed CPUE and the corresponding predicted one based on neural networks. There values of correlation coefficient by neural networks are rather better than those by MCMC based on the multiple imputation method. We suggested the simple way for factorial experiment to extract year trends based on the estimated CPUE by neural networks. It was found that the extracted CPUE year trend based on this simple way using neural networks is rather similar to that by generalized linear models assuming that there exists some fish in the missing cells as well as surrounding areas. The results are consistent with so-called constant square hypothesis using generalize linear models in the CPUE analyses for southern bluefin tuna.
計量生物学 27(1), 35-53, 2006-06-30
The Biometric Society of Japan