Prediction of zooplankton distribution in Southern ocean using machine learning

Krivoguz Denis Olegovich, Malko Sergey Vladimirovich, Semenova Anna Jurievna

The article discusses applying possibilities of machine learning to predict zooplankton concentrations in the Southern Ocean, depending on various environmental factors. Since zooplankton is an important part of the trophic chains of aquatic ecosystems, the role of predicting its quantity and distribution becomes more significant. Zooplankton of the Southern Ocean is represented by four large groups - copepods, euphausiids, salps and pteropods. They role in ecosystem of the Southern Ocean is to transfer energy from phytoplankton to the fishes. To reach our goals we analyzed modern approaches in zooplanktons modeling and identified 3 main groups of them - ecosystem, biogeochemical and size-based models. On opposite side, machine learning methods have a number of significant advantages, allowing you to make a prediction with a rather limited variety of ecosystem data.In this research we used data obtained from long-term monitoring mission of zooplankton in the Southern Ocean that was divided into test and training sets by 7 to 3 ratio. The spatial distribution of samples is mainly related to the Pacific sector of the Southern Ocean. Basic descriptive statistical analysis showed that high concentrations of zooplankton are common only in some areas of the Southern Ocean. Correlation analysis revealed no connection between analyzed environmental factors and the concentration of zooplankton, which indicates either the dependence of these concentrations on other factors not used in the study, or on their combined effect. To compare the possibilities of predicting zooplankton concentrations by machine learning, we chose 4 algorithms (k-nearest neighbors, random forest, AdaBoost, and artificial neural network) that were sequentially trained by the training data and then applied on test data. As a result, Random forest and AdaBoost algorithms showed the highest rating with 100% accuracy results. The worst result of prediction was shown by artificial neural networks with an accuracy of 86%.

Keywords:zooplankton, machine learning, Southern Ocean, modeling, statistical analysis