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 TOC concentrations and stocks at a 5 x 5 arc minute grid scale, using a deep learning model — a novel machine learning approach based on a new compilation of over 22,000 global TOC measurements and a new set of predictors, such as seafloor lithologies, grain size distribution, and an alpha-chlorophyll satellite data. In our study, we compared the predictions and discuss the limitations from various machine learning methods. Our findings reveal that the neural network approach outperforms methods such as k Nearest Neighbors and random forests, which tend to overfit to the training data, especially in highly heterogeneous and  complex geological settings. We provide estimates of mean TOC concentrations and total carbon stock in both continental  shelves and deep sea settings across various marine regions and oceans. Our model suggests that the upper 10 cm of oceanic sediments harbors approximately 171 Pg of TOC stock and has a mean TOC concentration of 0.68 %. Furthermore, we introduce a standardized methodology for quantifying predictive uncertainty using Monte Carlo dropout and present a map of information gain, that measures the expected increase in model knowledge achieved through in-situ sampling at specific locations which is  pivotal for sampling strategy planning.