Evaluating Sampling Bias and Model Uncertainty in Species Distribution Models of Marine Plankton Using Virtual Ecosystem Data

Zhibo Shao , B. B. Cael, Thelma Panaïotis, Anna Hickman, Stephanie Dutkiewicz, Ben Ward

Understanding the biodiversity and biogeography of plankton in the ocean is essential for predicting responses to environmental changes and informing ocean conservation and management strategies. Species distribution models (SDMs) are a pivotal tool in this regard. This study used data from a global marine ecosystem model as a testbed to assess the reliability of various SDMs, including Generalized Linear Model (GLM), Generalized Additive Model (GAM), Random Forest (RF), Boosted Regression Trees (BRT) and Artificial Neural Network (ANN). We used artificial datasets to replicate the sampling patterns of three datasets: a compiled dataset of global scope, the Tara Ocean dataset, and the Atlantic Meridional Transect (AMT) project. Our findings indicate that tree-based algorithms, RF and BRT, exhibit better predictive accuracy and stability compared to GLM, GAM, and ANN, especially when trained with more spatially resolved datasets. We highlight the significant influence of sampling bias on model performance, with models trained on more comprehensive global datasets outperforming those trained on more latitudinally and longitudinally biased data respectively (Tara and AMT). Furthermore, we demonstrate that broad spatial coverage is a more critical determinant of predictive skill than sample size alone, as simply increasing sampling density within a biased region is insufficient to overcome poor spatial representation. Overall, this research underscores the necessity of careful consideration of sampling strategies and model selection in plankton species distribution modelling.