A Machine Learning Framework for Assessing the Sensitivity of Regional Ocean Productivity to Climate Change

Net primary production (NPP) in the ocean is fundamental to marine food webs, supporting oxygen production for heterotrophic respiration and contributing to the long-term sequestration of carbon. Rising ocean temperatures associated with climate change are expected to alter NPP dynamics. However, it remains challenging to understand how different abiotic (especially sea temperature) and biotic factors influence marine NPP due to the complex network of interactions between these factors. This study introduces a flexible machine-learning-based framework for evaluating the sensitivity of NPP to variations in key environmental drivers, particularly sea temperature, by testing and comparing alternative machine learning algorithms. In the case study presented here, Support Vector Machines (SVM) achieved the highest predictive performance among the evaluated models. Variable-importance analysis of the best-performing algorithm, within the scope of this comparative framework, revealed that variables intrinsically linked to NPP, such as chlorophyll-a and solar radiation, play a key role in determining the predictive ability of the models. Meanwhile, sea temperature emerged as the key external factor influencing the performance of the models. The NPP exhibits a correlative sensitivity to increase of 1 °C in sea temperature, with relative changes ranging between 3% and 16%. These projections reflect model-based sensitivities derived from historical co-variation. Therefore, the results represent conditional projections under observed relationships. Although SVM performed best for this case study, the proposed framework is adaptable and can incorporate alternative algorithms, predictor sets and preprocessing strategies, enabling robust and transferable assessments of the sensitivity of regional ocean productivity to climate change.