NN-TOC v1: global prediction of total organic carbon in marine sediments using deep neural networks

Abstract. Spatial predictions of total organic carbon (TOC) concentrations and stocks are crucial for understanding marine sediments’ role as a significant carbon sink in the global carbon cycle. In this study, we present a geospatial prediction of global TOC concentrations and stocks on a 5 × 5 arcmin grid, using a novel neural network approach. We also provide and apply a new compilation of over 21 000 global TOC measurements and a new set of predictors, including features such as seafloor lithologies, benthic oxygen fluxes, and chlorophyll-a satellite data. Moreover, we compare different machine learning models based on their performance metrics and predictions and assess their strengths and limitations. For the dataset used, we find that the performance metrics of the models are comparable and that the neural network approach outperforms, on unseen data, methods such as k-nearest neighbours and random forests, which tend to overfit the training data. We provide estimates of mean TOC concentrations and stocks, both on continental shelves and in deep-sea settings across various marine regions and oceans. Our model suggests that the upper 10 cm of oceanic sediments harbour approximately 156 Pg of TOC stocks and have a mean TOC concentration of 0.61 %. Furthermore, we introduce a standardized methodology for quantifying predictive uncertainty using Monte Carlo dropout. The method was applied to our neural network model and underlying features to generate a map of information gain that measures the expected increase in model knowledge, achieved through additional sampling at specific locations, which is pivotal for sampling strategy planning.