Interannual inversions with continuous data and recent inversions with the CSIRO CC model Rachel Law, CSIRO Atmospheric Research Another set of pseudodata.

Interannual inversions with continuous data

and

recent inversions with the CSIRO CC model

Rachel Law, CSIRO Atmospheric Research

• Another set of pseudodata inversions (Law, 2004, ACP)

• Setting up to use real data

Changing models (to use ‘real’ winds)

Redefining regions

Testing with monthly data

Trying to add continuous data

ASPENDALE selected CO2

340

360

380

400

420

440

28-Jun 18-Jul 7-Aug 27-Aug 16-Sep 6-Oct 26-Oct

0

50

100

Cape Grim CO2

Aspendale CO2

Asp WS(m/s)

Asp wd/50+10

ASPENDALE selected CO2ws>2.5 m/s

355

365

375

385

395

2-Aug 4-Aug 6-Aug 8-Aug 10-Aug 12-Aug 14-Aug 16-Aug 18-Aug 20-Aug

0

50

100

Cape Grim CO2

Aspendale CO2

Asp WS(m/s)

Asp wd/50+10

Continuous data at Cape Grim (purple) and Aspendale (blue)

Wind speed > 2.5 m/s

BUT problem of size

T3-IAV base case (78 sites, solve 86-02): nsrc=4557, ndat=14040

116 region inversion, 35 sites with 4hr data, if solved over 86-02: nsrc=23665, ndat=1149750

AIM: Perform a T3L3 type inversion using pseudodata generated from interannually-varying fluxes (Friedlingstein land, LeQuere ocean) and compare results using monthly and 4 hourly data at 35 sites for 1982-1997

SOLUTION: Solve in 2 year overlapping segments adding 1 new year of data each time. Predicted sources from previous step used as priors in next step.

SHORT-CUT: Full covariance matrix not kept from one step to the next (but tests showed this wasn’t a problem)

Two example regions

Comparison with correct sources (black) and between 3 year inversion (red) and 1 year inversions (green, blue)

Difference in 1982-1997 mean source from correct source in gCm-2yr-1

(yellow/green good, blue/red bad)

Using 4 hr data

Some large biases over land regions – no significant improvement over monthly data

Using monthly data

Root mean square bias in the mean seasonal cycle in gCm-2yr-1

(blue good, red bad)

Using 4 hr data

Regions with sites nearby generally better than with monthly data

Using monthly data

Correlation of interannual variability between estimated and correct (red good, blue bad)

Using 4 hr data

Higher correlations over more regions (53 > 0.6) than with monthly data (17 > 0.6). Some low correlations over ocean – leakage of land signal.

Using monthly data

Magnitude of interannual variability compared to correct (red variability too large, blue variability too small)

Using 4 hr data

Variability too small or too large away from sites

Using monthly data

Magnitude of variability underestimated for land regions away from sites

CSIRO Conformal-Cubic Atmospheric Model (CCAM)

Stand-alone GCM or nudged with NCEP winds.

Uniform or stretched grid. Currently using ~200 km.

Trace gas transport using prescribed sources and on-line biospheric CO2.

Inversion set-up: 94 land regions, (42 also day-time flux), 46 (48) ocean regions.

6 month responses (hourly resolution) currently run for 1995, 1996, 1997.

Inversion using 1992-2002 monthly data only – 88 sites (81 GV, 7 WDCGG), data uncertainty: 1.5*rsd

Prior source: Fos95+CASA NEP+Taka (but no presub fields used)

Prior source uncertainty: ocean, area*150 gCm-2yr-1 land, tree area*1500+grass area*800+desert area*100 gCm-2yr-1

Total land – green

Total ocean – blue

Solid CCAM, dashed TC mean

Australian region source (12 month running mean) compared to TransCom mean

Note: CCAM case is sum of 16 regions, TransCom is one region and includes New Zealand. TC sources are sensitive to Baring Head

CCAM NZ source

Standard deviation of interannual component of fluxes in gCm-2yr-1

Prior source uncertainty in gCm-2yr-1

Adding continuous data

A small first step: use estimated sources and uncertainties from inversion with monthly data as prior sources for this inversion. Run inversion for 1 site (Cape Grim, 1997 hourly data).

What data uncertainty is appropriate?

Blue – obs green – unc 1.4 ppm red – unc 28 ppm

Change in 1997 mean source and uncertainty estimates with data uncertainty for four example regions

Blue – SE coast, mainland Australia Red – ocean to SW of Cape Grim

Blue – Western Brazil Red – Western Siberia

Ratio of estimated to prior uncertainty (for 1997 mean source) due to addition of Cape Grim hourly data with 7 ppm data uncertainty

Tasmania source estimate

Blue – original prior source, red – source from inversion of monthly data, green – set of inversions using hourly CGA with different data uncertainty, pale blue – smallest data uncertainty (1.4 ppm)

Source uncertainty shown by dashed line, mid-range (14ppm) case shown for set of inversions

Monthly data added little information about Tasmanian source, CGA data tends to weaken seasonal cycle. Note some months much better observed (e.g. Feb) than others (e.g. May). Larger shifts from prior in better observed months.

SE Australia source estimate

Blue – prior, red – monthly inversion, green - CGA hourly data

Monthly data moves sources closer to zero, reduces spring uptake, but only small reduction in uncertainty

Hourly data reduces seasonality further, tendency to shift uptake from Nov to Sep-Oct. Substantial decrease in uncertainty.

Ocean to south west of Cape Grim

Blue – prior, red – monthly inversion, green – hourly CGA data

Monthly data put large seasonality into region compared to Takahashi prior

Hourly data tends to moderate seasonality and make more consistent with Takahashi

Australian continental source

Red – monthly inversion, blue – addition of hourly Cape Grim Thick dashed line – 12 month running mean

Lose spring uptake and consequently shift annual mean. Result seems unlikely.

Future plans

• ‘Clean up’ method where possible e.g. monthly and continuous data in same step

• Use full covariance matrix in 2nd step

• Add more continuous sites

• Extend number of years

• Think about how to assess results – how do we know what’s a good result?

• Test temporal averaging of data

• Are there any ways to deal with transport model error?

Future plans – part 2