Similar to Pickers, Penelope: Novel quantification of regional fossil fuel CO2 reductions during COVID-19 lockdowns using atmospheric oxygen measurements
Similar to Pickers, Penelope: Novel quantification of regional fossil fuel CO2 reductions during COVID-19 lockdowns using atmospheric oxygen measurements (20)
Pickers, Penelope: Novel quantification of regional fossil fuel CO2 reductions during COVID-19 lockdowns using atmospheric oxygen measurements
1. Penelope A. Pickers, Andrew C. Manning, Corinne Le Quéré, Grant L. Forster, Ingrid T. Luijkx, Christoph Gerbig,
Leigh S. Fleming, William T. Sturges, Yasunori Tohjima, Karina Adcock, Tim Arnold and Caroline Dylag.
p.pickers@uea.ac.uk
Presented at the ICOS Science Conference, Utrecht, September 2022
Novel quantification of regional fossil fuel CO2 reductions during
COVID-19 lockdowns using atmospheric oxygen measurements
2. Coronavirus causes record fall in fossil-fuel emissions in 2020
GCB 2020 – Four bottom-up approaches:
1. UEA
2. Priestley
3. Carbon Monitor
4. GCB
Large range between different estimates.
Projections required (i.e. not real-time data).
Uncertainties inherent in methods and projections
not included.
“Despite the critical importance of CO2 emissions for
understanding global climate change, systems are not
in place to monitor global emissions in real time.”
– Le Quere et al. 2020
3. “Bottom-up” inventories lag real-time ~12-24 months
real-time mobility data helps, with higher
uncertainties
“Top-down” atmospheric inverse methods lack data
lag in modelling
no perfect ffCO2 tracer: radiocarbon (14CO2) is
precise but not currently continuous and Carbon
Monoxide (CO) is notoriously imprecise.
Wenger et al., ACP (2019)
Grid cell relative variability: ± 91-94%
Why is it difficult to quantify ffCO2 emissions quickly?
𝒇𝒇𝑪𝑶𝟐 𝑪𝑶 =
𝑪𝑶𝒎𝒆𝒂𝒔𝒖𝒓𝒆𝒅 − 𝑪𝑶𝒃𝒂𝒄𝒌𝒈𝒓𝒐𝒖𝒏𝒅
𝑹𝑪𝑶
TNO data courtesy of Hugo Denier van de Gon and colleagues
RCO is the CO:CO2 ratio for fossil fuel combustion.
4. APO = O2 + 1.1 x CO2
Atmospheric Potential Oxygen (APO)
APO is the sum of atmospheric CO2 and O2 measurements
where 1.1 is the mean –O2:CO2 ratio of terrestrial
biosphere exchange.
APO is a tracer invariant to terrestrial biosphere O2
and CO2 exchange.
𝒇𝒇𝑪𝑶𝟐 𝑨𝑷𝑶 =
𝑨𝑷𝑶𝒎𝒆𝒂𝒔𝒖𝒓𝒆𝒅 − 𝑨𝑷𝑶𝒃𝒂𝒄𝒌𝒈𝒓𝒐𝒖𝒏𝒅
𝑹𝑨𝑷𝑶
RAPO is the APO:CO2 ratio for fossil fuel combustion
RAPO has a much smaller range than RCO
APO can be measured continuously
5. Initial studies suggest APO shows potential
ffCO2[14CO2] and ffCO2[APO] from the same flasks.
Collected in Tsukuba, Japan 2016-2017
Ocean influences?
Short-term ocean-related O2 variability small at most sites
In our study, ocean influences are incorporated into the baseline Pickers et al. in prep
CO
APO
14CO2
ffCO2 from STILT
ffCO
2
from
tracers
0 10 20 30
0
10
20
30
CHE report D4.4
6. APO from Weybourne Atmospheric Observatory
Located on north Norfolk coast in UK
Class 2 ICOS station
WMO GAW regional station
UK National Centre for Atmospheric Science
Measurement Facility >12 year record of continuous O2 and CO2 measurement
7. Detecting COVID-19 signals in APO-derived ffCO2 from WAO
Too much variability in atmospheric transport to see anything in cumulative ffCO2
Pickers et al. Science Advances, 2022.
8. • Used the ‘rmweather’ R package
• Multiple “boosted regression trees” to describe the response
between descriptor and explanatory variables.
• Iterative process:
• Observations are repeatedly split using binary algorithm into two
homologous groups, known as ‘branches’, until the ‘tree’ is fully
grown (node purity is achieved).
• each tree is grown on out-of-bag data to create a “forest” of
decorrelated trees that are less susceptible to over-fitting.
2020-2021
prediction
Detecting COVID-19 signals in APO-derived ffCO2 from WAO
Trained a model of 300 trees using 2010-2019 data and then
used it to predict the 2020 “counterfactual case”, i.e. what we
would have expected with no pandemic.
Pickers et al. Science Advances, 2022.
9. 10 independent variables used for training:
Met data: wind speed, wind direction, air temperature,
relative humidity, atmospheric pressure.
Temporal variables: day of the year, day of the week,
hour of the day.
Radon concentration (from Apr 2018 only)
24-hour long hourly back-trajectories from Hysplit
atmospheric transport model (via Openair R package),
clustered into 8 groups (with k-means clustering).
Dependent variable => ffCO2 derived from APO
Detecting COVID-19 signals in APO-derived ffCO2 from WAO
Met data was quality controlled against independent
(but co-located) measurements from the UK Met Office
Pickers et al. Science Advances, 2022.
10. Mean decrease in ffCO2[APO] of
0.7 ppm for 2020.
APO = -23% (-14 to -32%)
Bottom-up estimates:
• Carbon Monitor = -8%
• UEA = -21% (-13 to -30%)
• UK BEIS = -17%
(all reductions calculated for
01Feb2020‐31Jan2021)
Detecting COVID-19 signals in APO-derived ffCO2 from WAO
Second wave
First wave
Pickers et al. Science Advances, 2022.
Blue shading => ±2σ (95%) SD of the 2011–2019 mean (solid blue line)
Pink shading => uncertainty from imperfections in ML prediction
ffCO2[APO] uncertainties exist but cancel
11. Pickers et al. Science Advances, 2022.
Do we trust the results?
Machine learning performance looks good
12. Do we trust the results?
Pickers et al. Science Advances, 2022.
No substantial changes in RAPO over study period Signal holds for different prediction start dates
Signal pattern holds for training period of only two years (2011-2013 or 2017-2019) but
magnitude is reduced => longer dataset is helpful
13. • COVID-19 signal detectable using APO alone?
• COVID-19 signal detectable using APO
combined with machine learning?
• COVID-19 signal detectable using CO2
combined with machine learning?
• COVID-19 signal detectable using 14CO2?
COVID-19 signal summary
?
Pickers et al. Science Advances, 2022.
Signal emerges ~6 months after lockdown.
Cannot differentiate ffCO2 emissions reductions from
potential biospheric related reduction, such as late
summer 2020 UK heatwave
14. Next steps
• Long-term decline in ffCO2 visible in 2011-2019 pre-pandemic model-observation
data at Weybourne.
• Better baseline determination using radon data.
• More accurate RAPO values and better understanding of RAPO uncertainty
• 2nd UK site at Heathfield in Sussex
December 2020 CarbonBrief headline based on latest Global Carbon Budget analysis
24% for TNO, 17% for COFFEE. Linear scale!
The rmweather package utilizes Random Forest, an ensemble decision tree machine learning method that splits observations using a binary algorithm into two homologous groups, known as branches, repeating the process until the ‘tree’ is fully grown (node purity is achieved). Decision trees are prone to over‐fitting, but Random Forest mitigates this by growing many individual decision trees from a training set in a process called bagging (bootstrap aggregation), which creates a forest of decorrelated trees, since each has been grown on different subsets of the training set.
10 independent variables
Richard Betts: -1 ppm for -18% reduction in global emissions.
The estimates are not directly comparable because the ffCO2[APO] top‐down signal is in ppm and is not representative of the UK as a whole (WAO is influenced mainly by south westerly winds, see Supplementary Figure 8) while the bottom‐up COVID‐19 signals are in MtCO2 and are UK totals.
ffCO2[APO] reduction is relative to the counterfactual prediction for 2020, bottom‐up estimates are relative to emissions for 01March‐31Jul2019; uncertainty ranges not available for Carbon Monitor and UK BEIS estimates). Like ffCO2[APO], the UK BEIS national inventory also exhibits a reduction in ffCO2 emissions during January followed by a recovery in February, although the changes in UK BEIS are not as large.