Modelling global mesozooplankton biomass using machine learning

• Four machine learning algorithms (boosted regression trees, random forest, artificial neural network, and support vector machine) were used to model the global distribution of mesozooplankton biomass. • Random Forest (RF) achieved the best predictive accuracy with the highest R 2 and lowest RMSE . • Our model created a more complete global map of mesozooplankton biomass and reproduced their seasonal and spatial patterns, which is helpful for validating and optimizing process-based marine ecosystem models. • A scaling relationship between the mesozooplankton biomass and Chl a concentration was identified and can be used as an emergent constraint for model validation and development. • Our data-driven model, based on RF algorithms, predicts a 3% decline in global mesozooplankton biomass by the end of this century under the “business-as-usual” scenarios. This may result in a series of consequences, such as reduced fishery production. Mesozooplankton are a crucial link between primary producers and higher trophic levels and play a vital role in marine food webs, biological carbon pumps, and sustaining fishery resources. However, the global distribution of mesozooplankton biomass and the relevant controlling mechanisms remain elusive. We compared four machine learning algorithms (Boosted Regression Trees, Random Forest, Artificial Neural Network, and Support Vector Machine) to model the spatiotemporal distributions of global mesozooplankton biomass. These algorithms were trained on a compiled dataset of published mesozooplankton biomass observations with corresponding environmental predictors from contemporaneous satellite observations (temperature, chlorophyll, salinity, and mixed layer depth). We found that Random Forest achieved the best predictive accuracy with R 2 and RMSE (Root Mean Standard Error) of 0.57 and 0.39, respectively. Also, the global distribution of mesozooplankton biomass predicted by the Random Forest model was more consistent with the observational data than other models. We used the Random Forest model to create a global map of mesozooplankton biomass which serves as a reference for validating process-based ecosystem models. The model outputs confirm that environmental factors, especially surface Chl a , a proxy for prey availability, significantly correlate with the spatiotemporal distribution of mesozooplankton biomass. The scaling relationship between the mesozooplankton biomass and Chl a can be used as an emergent constraint for model validation and development. Moreover, our model predicts that the global total mesozooplankton biomass will decrease by 3% by the end of this century under the “business-as-usual” scenarios, potentially reducing fishery production and carbon sequestration. Our study contributes to predicting global mesozooplankton biomass and provides deep insights into the underlying environmental impacts on the distribution of mesozooplankton biomass.