Standardizing catch per unit effort by machine learning techniques in longline fisheries: a case study of bigeye tuna in the Atlantic Ocean

Support vector machine (SVM) is shown to have better performance in catch per unit of effort (CPUE) standardization than other methods. The SVM performance highly relates to its parameters selection and has not been discussed in CPUE standardization. Analyzing the influence of parameter selection on SVM performance for CPUE standardization could improve model construction and performance, and thus provide useful information to stock assessment and management. We applied SVM to standardize longline catch per unit fishing effort of fishery data for bigeye tuna (Thunnus obesus) in the tropical fishing area of Atlantic Ocean and evaluated three parameters optimization methods: a Grid Search method, and two improved hybrid algorithms, namely SVMs in combination with the particle swarm optimization (PSO-SVM), and genetic algorithms (GA-SVM), in order to increase the strength of SVM. The mean absolute error (MAE), mean square error (MSE), three types of correlation coefficients and the normalized mean square error (NMSE) were computed to compare the algorithm performances. The PSO-SVM and GA-SVM algorithms had particularly high performances of indicative values in the training data and dataset, and the performances of PSO-SVM were marginally better than GA-SVM. The Grid search algorithm had best performances of indicative values in testing data. In general, PSO was appropriate to optimize the SVM parameters in CPUE standardization. The standardized CPUE was unstable and low from 2007 to 2011, increased during 2011-2013, then decreased from 2015 to 2017. The abundance index was lower compared with before 2000 and showed a decreasing trend in recent years.