An Argo based estimate of Oxygen (O_2) at 150 m is presented for the Southern Ocean (SO) from T/S, O_2 Argo profiles collected during 2008-2012. The method is based on supervised machine learning, i.e. Random Forest (RF) regression, and provides an estimate for O_2 on gridded Argo T/S fields. Results show that the Southern Ocean State Estimate (SOSE) and the World Ocean Atlas 2013 climatology may overestimate annual mean O_2 in the SO, both on a global and basin scale. A large regional bias is found east of Argentina, where high O_2 values in the Argo based estimate are closer to the coast compared to other products. SOSE may also underestimate the annual cycle of O_2. Regions where the RF method does not perform well (e.g. eastern boundaries) are identified comparing the actual SOSE O_2 fields to the RF estimate from model profiles co-located with observations. The RF based method presented here has the potential to improve our understanding of O_2 annual mean fields and variability from available (sparse) O_2 measurements. Also, it may guide the design of future enhancements to the current array of O_2 profiling floats, and prove effective for other biogeochemical variables (e.g. nutrients and carbon).

1. Estimating Oxygen in the
Southern Ocean using Argo
Temperature and Salinity
Donata Giglio
Vyacheslav Lyubchich
Matthew R. Mazloff

2. Motivation
Can we improve mapping of sparse oxygen (O2) observations
using the relatively numerous Argo T/S profiles?
Number of profiles in 1×1 degree bins during June, July, August.
Argo O2 Argo T/S
during 2008-2012

3. Motivation
Can we improve mapping of sparse oxygen (O2) observations
using the relatively numerous Argo T/S profiles?
Number of profiles in 1×1 degree bins during June, July, August.
WOA13 O2 Argo O2 Argo T/S
since early 1900s during 2008-2012

4. Outline
• Data
• Method
• Intro
• Diagnostics
• Assessment
• Results
• Time mean
• Annual variability
• Simulating different sampling scenarios
• Summary
• Ongoing work

5. Data at 150 m
Argo profiles: calibrated O2 (Drucker and Riser, 2016), T/S
2008-2012
SOSE Biogeochemical Southern Ocean State
Estimate O2, T/S
WOA13 World Ocean Atlas 2013 O2 since early 1900s
· Argo O2 profiles

6. Method: outline
1. Training a machine learning method to predict O2 knowing T/S, using
A. Argo profiles
B. SOSE profiles
2. Predicting O2 using
A. Argo T/S (then gridding O2 predictions)
B. gridded SOSE T/S
3. For B., comparing O2 predictions to actual O2 SOSE fields
4. Comparing Argo O2 RF based estimates to other products (SOSE, WOA13)

7. Random Forest Regression: intro
(modeling O2 with Temperature, Salinity, Month, Year, Latitude, Longitude)
sparse observations a large dataset available
Basic element of a random forest regression is a regression tree.

8. Random Forest Regression: intro
(modeling O2 with Temperature, Salinity, Month, Year, Latitude, Longitude)
sparse observations a large dataset available
Basic element of a random forest regression is a regression tree.
S >= 34.4
lat < −58
T >= −0.29
T >= 0.88
lat >= −36
S >= 34.1
T >= −0.78
lat >= −39
T < 2
T >= 0.52
T >= 8
256
100%
238
60%
221
18%
207
13%
196
8%
221
5%
265
5%
245
42%
229
15%
253
27%
282
40%
268
23%
265
21%
239
2%
268
19%
256
8%
243
4%
269
4%
277
11%
302
2%
301
17%
237
1%
303
16%
yes no
Random Forest
=
Many Trees

9. Random Forest Regression: diagnostics
(modeling O2 with Temperature, Salinity, Month, Year, Latitude, Longitude)
MSE (mean squared error)
sample data
used to build tree n
Out-of-bag data used to compute:
• the MSE for tree n
•
0 100 200 300 400 500
0
20
40
60
80
100
120
Number of trees
Meansquarederror,(µmolkg−1
)2
Data source
Argo profiles
SOSE profiles
T T T T
O2 O2 O2 O2
S S S S
T T
O2 O2
S S

10. Random Forest Regression: diagnostics
(modeling O2 with Temperature, Salinity, Month, Year, Latitude, Longitude)
Permutation-based importance of predictors
sample data
used to build tree n
Out-of-bag data used to compute:
• the MSE for tree n
• the importance of predictor i as difference between MSE using vs .
e.g. for T
O2 O2_obs1, O2_obs2
T T1, T2 T2, T1
S S1, S2 S1, S2
lat lat1, lat2 lat1, lat2
lon lon1, lon2 lon1, lon2
month m1, m2 m1, m2
year yr1, yr2 yr1, yr2

11. Random Forest Regression: diagnostics
(modeling O2 with Temperature, Salinity, Month, Year, Latitude, Longitude)
Permutation-based importance of predictors
sample data
used to build tree n
Out-of-bag data used to compute:
• the MSE for tree n
• the importance of predictor i as difference between MSE using vs .
e.g. for T
O2 O2_obs1, O2_obs2
T T1, T2 T2, T1
S S1, S2 S1, S2
lat lat1, lat2 lat1, lat2
lon lon1, lon2 lon1, lon2
month m1, m2 m1, m2
year yr1, yr2 yr1, yr2
S T lat lon Year Month
Data source
Argo profiles
SOSE profiles
0
500
1000
1500
Ii,(µmolkg−1
)2

12. Outline
• Data
• Method
• Intro
• Diagnostics
• Assessment
• Results
• Time mean
• Annual variability
• Simulating different sampling scenarios
• Summary
• Ongoing work

13. Random Forest Regression: assessment
(modeling O2 with Temperature, Salinity, Month, Year, Latitude, Longitude)
annual mean O2 at 150 m
���
���
SOSE O2

14. Random Forest Regression: assessment
(modeling O2 with Temperature, Salinity, Month, Year, Latitude, Longitude)
���
���
SOSE O2
· Argo O2
profiles
annual mean O2 at 150 m

15. Random Forest Regression: assessment
(modeling O2 with Temperature, Salinity, Month, Year, Latitude, Longitude)
• < 1.5% error in 65% of the domain
• Large method bias near eastern boundaries
���
���
O2 RF estimate minus SOSE O2
— 1.5% diff
···· 3% diff
annual mean O2 at 150 m

16. Random Forest Regression: assessment
(modeling O2 with Temperature, Salinity, Month, Year, Latitude, Longitude)
• < 1.5% error in 65% of the domain
• Large method bias near eastern boundaries
���
���
������O2 RF estimate minus SOSE O2 SOSE σO2
— 1.5% diff
···· 3% diff
annual mean O2 at 150 m

17. Outline
• Data
• Method
• Intro
• Diagnostics
• Assessment
• Results
• Time mean
• Annual variability
• Simulating different sampling scenarios
• Summary
• Ongoing work

18. SOSE and WOA13 overestimate annual mean O2
������
SOSE O2 minus RF Argo O2 WOA13 minus RF Argo O2
Annual mean O2 at 150 m

19. ���
���
AnnualmeanO2,
µmolkg-1
Annual O2 anomaly at 150 m
• WOA13: sparse observations, especially in winter
• is SOSE underestimating annual O2 variability ?
RF Argo
RF SOSE (—)
SOSE (- -)
WOA13

20. Outline
• Data
• Method
• Intro
• Diagnostics
• Assessment
• Results
• Time mean
• Annual variability
• Simulating different sampling scenarios
• Summary
• Ongoing work

21. Simulating different sampling scenarios
· Argo O2
profiles
oversampling in eastern boundary regions
increases errors in the ocean interior

22. Summary
An extensive array of O2 profiling floats is the long-term goal for the scientific
community.
• RF framework and Argo T/S improve understanding of O2 annual mean and
variability at present, identifying biases in SOSE and WOA13.
• Identified O2 biases in SOSE and WOA13 are consistent with T/S differences
from Argo.
• Using sea surface height and surface chlorophyll as additional predictors in RF
gives similar results, as T and S remain the most informative predictors.
• A sampling representing all statistical regimes is necessary for RF skill. We find
current under-sampling in eastern boundary regions leads to large errors in these
areas, while oversampling these regions increases errors in the interior ocean.
Giglio, Donata, Lyubchich V., and Mazloff M. R.: Estimating Oxygen in the Southern Ocean
using Argo Temperature and Salinity, Journal of Geophysical Research: Oceans, accepted

23. Summary
An extensive array of O2 profiling floats is the long-term goal for the scientific community.
• RF framework and Argo T/S improve understanding of O2 annual mean and variability at present, identifying biases in SOSE
and WOA13.
• Identified O2 biases in SOSE and WOA13 are consistent with T/S differences from Argo.
• Using sea surface height and surface chlorophyll as additional predictors in RF gives similar results, as T and S remain the
most informative predictors.
• A sampling representing all statistical regimes is necessary for RF skill. We find current under-sampling in eastern boundary
regions leads to large errors in these areas, while oversampling these regions increases errors in the interior ocean.
Giglio, Donata, Lyubchich V., and Mazloff M. R.: Estimating Oxygen in the Southern Ocean
using Argo Temperature and Salinity, Journal of Geophysical Research: Oceans, accepted
Ongoing work
• Using Argo data together with other available observations
• Comparison with previous methods
• Comparison with new methods (e.g. NN, spatial NN/RF)
(µmol kg-1)2
RF NN spatial RF spatial NN
MSE 43 50 57.3 89
courtesy of Huang Huang (SAMSI, Duke University)

25. Annual mean O2 at 150 m
��� ���
��� ���
������
��
O2, µmol kg-1
Global zonal
average
RF Argo
SOSE (- -)
WOA13
SOSE and WOA13 overestimate annual mean O2