Natural mortality estimation using tree-based ensemble learning methods

Empirical methods have been widely used in estimating fish natural mortality rate (M). However, these empirical methods are often perceived as being less reliable than direct methods. To improve the predictive performance and reliability of empirical methods, we develop ensemble learning models, including bagging trees, random forests, and boosting trees, to predict M based on a dataset with 324 stocks of both Chondrichthyes and Osteichthyes. Three common life-history parameters are used as predictors: the maximum age and two growth parameters (growth coefficient and asymptotic length). In addition, taxonomic rank class is included to distinguish elasmobranchs and teleosts. The results indicate that the tree-based ensemble learning methods significantly improve the accuracy of M estimate, compared to the traditional linear regression models and basic regression tree models. Among ensemble learning methods, the boosting trees and random forests perform best on training dataset, but the former performs a slightly better on test dataset. We develop four boosting tree models for estimating M based on varying life-history parameters and construct an R package for M estimation. The ensemble learning methods used in this paper can also be adopted in examining correlations among other life-history parameters and fisheries assessments to improve prediction accuracy.

This is a collaborative research project supervised by Professor Shijie Zhou in CSIRO and Professor Yougan Wang in QUT.