A Machine Learning (kNN) Approach to Predicting Global Seafloor Total Organic Carbon. Issue 1 (23rd January 2019)

• Record Type: Journal Article
• Authors: Lee, Taylor R.; Wood, Warren T.; Phrampus, Benjamin J.
• Is Part Of: Global biogeochemical cycles. Volume 33:Issue 1(2019:Jan.)
• Page Start: 37
• Page End: 46
• Publication Date: 2019-01-23
• Subjects: machine learning -- total organic carbon -- seafloor properties -- interpolation techniques -- global prediction; Biogeochemical cycles -- Periodicals; Electronic journals; 577.1405
• DOI: 10.1029/2018GB005992
• Languages: English
• ISSNs: 0886-6236

Abstract

Abstract: Seafloor properties, including total organic carbon (TOC), are sparsely measured on a global scale, and interpolation (prediction) techniques are often used as a proxy for observation. Previous geospatial interpolations of seafloor TOC exhibit gaps where little to no observed data exists. In contrast, recent machine learning techniques, relying on geophysical and geochemical properties (e.g., seafloor biomass, porosity, and distance from coast), show promise in making comprehensive, statistically optimal predictions. Here we apply a nonparametric (i.e., data‐driven) machine learning algorithm, specifically k‐nearest neighbors (kNN), to estimate the global distribution of seafloor TOC. Our results include predictor (feature) selection specifically designed to mitigate bias and produce a statistically optimal estimation of seafloor TOC, with uncertainty, at 5 × 5‐arc minute resolution. Analysis of parameter space sample density provides a guide for future sampling. One use for this prediction is to constrain a global inventory, indicating that just the upper 5 cm of the seafloor contains about 87 ± 43 gigatons of carbon (Gt C) in organic form. Key Points: Total organic carbon in the shallow seafloor is a fundamental quantity for many subsurface processes but is only very sparsely sampled We use machine learning techniques to predict total organic carbon for the entire seafloor, with uncertainty, a 5 × 5‐arc minute grid Parameter space proximity indicates where and what kinds of future measurements are the most optimal for reducing prediction uncertainty