Application of the random forest algorithm for mapping potential fishing zones of Rastrelliger kanagurta off the east coast of peninsular Malaysia

Abstract

Indian mackerel, Rastrelliger kanagurta, play an important role in the marine fisheries of Malaysia, and their distribution is reported to be influenced by various oceanographic conditions. In this study, a modeling framework of machine learning techniques (random forest [RF]) and spatial analysis with geographic information systems (hotspot analysis) was used to predict and map potential fishing zones of R. kanagurta in the exclusive economic zone (EEZ) waters off the east coast of peninsular Malaysia. The RF model was constructed with a training dataset (n = 2535) and externally validated by a testing dataset (n = 1087). Model validation using Lin’s concordance correlation coefficient (CCC) on the testing dataset indicated good agreement (CCC = 0.811) between the predicted and observed catch per unit effort (CPUE) of R. kanagurta. The variable importance plot revealed the relative importance of four predictors of potential habitat, ranked in decreasing order of importance: sea surface height anomaly (SSHA), eddy kinetic energy (EKE), sea surface temperature (SST) and surface chlorophyll-a (CHL) concentration. Partial dependence shows that R. kanagurta prefers a habitat with the following parameters: SSHA -0.05–0.20 m, EKE < 0.02 m2/s2 or > 0.07 m2/s2, SST 30–31 °C and CHL 0.1–0.3 mg/m3. Spatial distribution of predicted CPUE hotspots were found to be closely related to the thermal fronts, which influenced by the monsoonal wind and occurrence of upwelling event in the study area. Our study indicates that cost-effective satellite-derived products can be used to predict potential fishing zones of R. kanagurta and provide useful information on relationship between environmental factors and CPUE of R. kanagurta.