Predicting Global Seafloor Organic Carbon Burial Rates: A Deep Learning Approach with Uncertainty Quantification

Sediment accumulation rate is recognized as the primary parameter influencing the burial rate of organic carbon and other compounds in marine sediments. The prediction of a global map for burial rates is challenging due to the limited availability of measurements for total organic carbon (TOC) and sediment accumulation rates from the seafloor. Recent advancements in machine learning, including techniques such as K nearest Neighbours and Random Forests, have demonstrated promise in producing comprehensive predictions utilizing global maps of oceanic properties.

 

In this study, we introduce a sophisticated approach based on a newly developed deep neural network (DNN) model tailored for geospatial predictions. Employing few-shot learning techniques, such as the incorporation of prior physical knowledge into the model, along with strategies like multi-task learning and semi-supervised learning, enhances predictions amidst sparse data availability. Moreover, p​​​​​redictions of the global distribution of seafloor TOC and sediment accumulation rates presented here are coupled with uncertainty maps computed using Monte Carlo Dropout, a Bayesian approximation method that effectively inform about the degree of the model predictibility. With our results, we not only explore the global distribution of burial rates of organic carbon but also offer insights into the global carbon stocks in various marine regions.