NDP-101 (2019)

NCEI Accession 0206205 Data and Documentation Files

Global surface-ocean partial pressure of carbon dioxide (pCO₂) estimates from a machine learning ensemble: CSIR-ML6 v2019a (NCEI Accession 0206205)

Luke Gregor^{1, 2, 3}, Alice D. Lebehot^1,2, Schalk Kok⁴, Pedro M. Scheel Monteiro¹

Prepared by Alex Kozyr⁵

¹SOCCO, Council for Scientific and Industrial Research, Cape Town, 7700, South Africa
²MaRe, Marine Research Institute, University of Cape Town, Cape Town, 7700, South Africa
³Environmental Physics, Institute of Biogeochemistry and Pollutant Dynamics, ETH Zürich, 8092 Zürich, Switzerland
⁴Department of Mechanical & Aeronautical Engineering, University of Pretoria, Pretoria, 0028, South Africa
⁵National Centers for Environmental Information, National Oceanic and Atmospheric Administration, 1315 East-West Highway, Silver Spring, MD 20910-3282

Abstract

This NCEI accession contains surface-ocean partial pressure of carbon dioxide (pCO₂) that the ensemble mean of six two-step clustering-regression machine learning methods. The ensemble is a combination of two clustering approaches and three regression methods. For the clustering approaches, we use K-means clustering (21 clusters) and open ocean CO₂ biomes as defined by Fay and McKinley (2014). Three machine learning regression methods are applied to each of these two clustering methods. These machine learning methods are feed-forward neural-network (FFN), support vector regression (SVR) and gradient boosted machine using decision trees (GBM). The final estimate of surface ocean pCO₂ is the average of the six machine learning estimates resulting in a monthly by 1° ⨉ 1° resolution product that extends from the start of 1982 to the end of 2016. Sea-air fluxes (FCO₂) calculated from pCO₂ are also presented in the data. The discrete boundaries of the clustering approach result in semi-discrete discontinuities in pCO₂ and fCO₂ estimates. These are smoothed by applying a 3 ⨉ 3 ⨉ 3 convolution (moving average) to the dataset in time, latitude and longitude.

Method citation:

Gregor, L., Lebehot, A. D., Kok, S., and Scheel Monteiro, P. M.: A comparative assessment of the uncertainties of global surface­-ocean CO2 estimates using a machine learning ensemble (CSIR-ML6 version 2019a) – have we hit the wall?, Geosci. Model Dev. Discuss., https://doi.org/10.5194/gmd-2019-46, in review, 2019.

Variables:

COORDINATES
time (420)
units = seconds since 2000-01-01
min = 1982-01-15
max = 2016-12-15
step = month
lat (180)
	long_name = latitude
	units = degrees_north
	min = -89.5
	max = 89.5
	step = 1.0
lon (360)
	long_name = longitude
	units = degrees_east
	min = -180
	max = 180
	step = 1.0 
	
VARIABLES
pCO2sea_raw  (time, lat, lon) 
	long_name = partial pressure of surface ocean CO2
        units = µatm
        description = The ensemble mean of six machine learning methods that first cluster data and then apply regression to the clusters.


pCO2sea_smooth (time, lat, lon) 
	long_name = smoothed partial pressure of surface ocean CO2
	units = µatm
        description = The discrete boundaries of the clustering approach result in semi-discrete discontinuities in pCO2 and FCO2 estimates. These are smoothed by applying a 3 ⨉ 3 ⨉ 3 convolution (moving average) to the dataset in time, latitude and longitude. 


FCO2_raw (time, lat, lon)
        long_name = sea-air CO2 flux 
        units = molC/m2/yr
        Description = sea-air CO2 flux calculated after Landschutzer et al (2016) as the product of the following: (pCO_2^sea-pCO_2^air) where positive values are outgassing from sea to air; the Wanninkhof (1992) parameterisation of k_w scaled globally to 16 cm/hr with wind from ERA-interim (Dee et al, 2011); K_0 from Weiss (1974) using the OISSTv2 sea surface temperature product (Reynolds et al. 2007) and EN4 sea salinity (Good et al. 2013); (1 - ice_frac) as described in Butterworth and Miller (2016).


FCO2_smooth (time, lat, lon)
        long_name = smoothed sea-air CO2 flux
        units = molC/m2/yr
        description = The discrete boundaries of the clustering approach result in semi-discrete discontinuities in pCO2 and FCO2 estimates. These are smoothed by applying a 3 ⨉ 3 ⨉ 3 convolution (moving average) to the dataset in time, latitude and longitude. 


pCO2air (time, lat, lon) 
        long_name = partial pressure of atmospheric CO2
	units = µatm
	description = Atmospheric pCO2 from CarboScope v1.7 (Rodenbeck et al 2014)
	source = http://www.bgc-jena.mpg.de/CarboScope/ 


seamask (lat, lon) 
	description = boolean mask where True is ocean and False is land or NULL
	source = www.nodc.noaa.gov/ocads/oceans/SPCO2_1982_2015_ETH_SOM_FFN.html
	
	

Please cite this data set as:
Gregor, Luke; Lebehot, Alice D.; Kok, Schalk; Monteiro, Pedro M. Scheel (2019). Global surface-ocean partial pressure of carbon dioxide (pCO₂) estimates from a machine learning ensemble: CSIR-ML6 v2019a (NCEI Accession 0206205). Version 1.1. NOAA National Centers for Environmental Information Dataset.