CO2 variability simulated with daily fluxes

1
CO2 variability simulated with daily fluxes

• Shamil Maksyutov, Misa Ishizawa
• Frontier Research System for Global Change,
  Yokohama Japan

Transcom workshop Tsukuba 2004
2
Should one simulate the high frequency
variability

• Pro
• Inverse models relying on snapshot observations
  (flask data) have to deal with observation noise
  translated directly into noise in fluxes
  (interannual or monthly). ? Perfect forward model
  simulation is supposed to reduce this source of
  noise
• Compared to monthly average data use, more
  uncertainty reduction can be expected with the
  same data, by applying different weight
  (observation error) to each measurement, avoiding
  conservative error estimate for aggregated data.
  Kind of data aggregation error or bias

3
Should one simulate the high frequency
variability
4
Should one simulate the high frequency
variability

• Contra
• Forward model simulation errors are even more
  evident in high frequency time series, including
• Transport model error (limited model resolution
  in time and space, numerical diffusion, imperfect
  physics, etc)
• Emission field errors
• Terrestrial ecosystem flux simulation with
  ecosystem models daily flux simulation often
  fails even at flux tower sites with well known
  hydrology, phenology, vegetation type, when
  global/regional gridded meteo data sets are used
  as forcing for process based models.

5
Multiyear simulation of atmospheric CO2
variability in global scale

• Atmospheric tracer transport NIES transport
  model with NCEP wind interpolated to 2.5 or 1
  degree resolution. Modification increased
  tropospheric mixing by prescribed turbulence.
• Fossil fuel emission (CDIAC)
• Oceanic flux (Takahashi, 1999, 2002), as in T3
  protocol
• Terrestrial biosphere
• a) CASA Randerson 1997 (Transcom)
• b) Biome BGC (Fujita et al 2003)

6
Biospheric model fluxes at daily resolution

• Biome-BGC model (S. Running, P. Thornton) v. 4.12
• Ecosystem type map (derived from Matthews by R.
  Hunt, 1996) 1x1 deg.
• 1x1 deg Zobler soil data set ( clay, sand, silt
  for water holding capacity simulation)
• 1.8 deg 6hourly NCEP reanalysis data set (tmin,
  tmax, temp, precip, short wave radiation)
  interpolated to 1x1 deg.
• 2 versions of SWR algorithm tried a) - Mtclim
  processor, b) NCEP reanalysis

7
Model vs. observations simulating spikes
8
Model vs. observations simulating spikes
9
Simulating continuous observations at Hateruma

• Observations hourly data interpolated to daily
  average. Hourly data show much larger variability
  as compared to model simulation
• Model simulation (6hourly output) performed with
  CASA (monthly) and Biome-BGC (daily) fluxes.
• Surprisingly, monthly CASA fluxes are sufficient
  in many cases for simulating the synoptic scale
  variability, while dailiy Biome-BGC fluxes do not
  have much advantage.

10
Short term variability at marine site Hateruma
11
Extracting short term variability

• Seasonal cycle fit curve is subtracted from both
  observations and simulations respectively.

12
Short term variability at marine site Hateruma
Comparison shows better correlation in winter
13
Short term variability at marine site Hateruma
Breakdown into components show anti-correlation
of biospheric and fossil fuel components in
summer may be a reason for difficulty to
simulate In winter similar magnitude and sign.
14
Short term variability at land site ITN
1997-98
15
Short term variability breakdown into
components ITN
In summer dominated by biospheric flux
contribution In winter biospheric and fossil
fuel are of similar order of magnitude
1997-98
16
Data selection issue for land site

• Preferable treatment is to select well mixed
  (usually afternoon condition)
• Selecting afternoon values at low towers may
  still lead to significant difference vs 500m
  level in winter
• For tall towers (500 m LEF, ITN) difference
  between daily and afternoon is small, compared to
  short term variability.

17
Short term variability breakdown into
components LEF
1995-96
18
Correlation as a measure of the forward model
performance (LEF)

• Better correlation in winter, CASA vs Biome BGC

19
Few formulas

• Objective function for optimization
• Data uncertainty in Globalview based analysis
  defined as observation error (variability plus
  analysis error including calibration offset
  uncertainty)
• In Bayesian inversion (as in Tarantola) flux is a
  linear function of (observation-model) mismatch,
  so it makes sense reducing mismatch at synoptic
  scale too. However posterior flux uncertainty is
  not influenced by the mismatch.

20
Summary

• Taking into account high frequency variability
  may improve inverse modeling performance, reduce
  actual flux noise.
• Even simple model setup with relatively crude
  resolution shows good correlation at time scales
  of 3-5 days most of the year
• Improvements in the inverse model theory are
  needed to incorporate model error and
  observation-model mismatch into the flux error
  estimate.