Comparison of machine learning models within different spatial resolutions for predicting the bigeye tuna fishing grounds in tropical waters of the Atlantic Ocean

Abstract

To understand the effects of the machine learning models and the spatial resolutions on the prediction accuracy of bigeye tuna (Thunnus obesus) fishing grounds, logbook data of 13 Chinese longliners operating in the high seas of the Atlantic Ocean from 2016 to 2019 were collected. The environmental factors were selected based on the correlation analysis of calculation of catch per unit effort (CPUE) and the marine vertical environmental factors. Five machine learning models: random forest, gradient-boosting decision tree, K-nearest neighbor, logistic regression and stacking ensemble learning (STK) within four spatial resolutions of .5° × .5°, 1° × 1°, 2° × 2° and 5° × 5° grids were constructed and compared. Results showed that (1) the prediction performance of STK model was the best, with the highest scores of the four evaluation indexes, accuracy (Acc), precision (P), recall (R), and F1-score (F1), and the highest correct prediction rate for predicting “high CPUE fishing ground”; (2) models within the spatial resolution of 1° × 1° grids predicted the better results compared with .5° × .5°, 2° × 2° and 5° × 5° grids; (3) the vertical environmental factors selected based on the correlation analysis could be used as reliable predictors in the models. Results suggested that using STK within 1° × 1° grids could improve the generalization performance and prediction accuracy for predicting the bigeye tuna fishing grounds in the Atlantic Ocean.