Predictive performance from abundance distribution models of Vinciguerria lucetia larvae in the southern portion of the California current system using XGBOOST

Abstract

Vinciguerria lucetia is a mesopelagic fish whose larvae show an almost permanent presence in the southern portion of the California Current System. Due to its sensitivity to environmental changes, the species has been considered an indicator of water masses and interannual variability. Fish larvae abundance registered from 1997 to 2015 by the program Investigaciones Mexicanas de la Corriente de California was used to predict the abundance distribution of V. lucetia larvae under two extreme thermal conditions (2000 La Niña and 2015 El Niño), utilizing the novel machine learning algorithm eXtreme Gradient Boosting (XGBOOST). The data were segmented into COLD and WARM groups based on the mean sea surface temperature recorded at each station and contrasted with an undivided TOTAL group. Models were generated using 12 environmental and biological predictor features. Root-mean-squared logarithm error (RMSLE) was used as a prediction performance metric for both internal and external validation. The COLD model showed the best performance for the internal validation with a lower RMSLE value, while the TOTAL model for both the coldest and warmest external validation presented the lowest RMSLE values. The external validation demonstrated models that accurately predicted the spatial distribution; however, none of the models were able to accurately predict the same abundance magnitude observed in both extreme thermal conditions. Nevertheless, XGBOOST shows promise for describing the future distribution traits of V. lucetia.