Looking at the climate changeecosystem problem: How do we gain scientific knowledge from data and mo

1
Looking at the climate changeecosystem problem
How do we gain scientific knowledge from data and
models at multiple temporal and spatial
scales? Steven C. Wofsy Prof. of Atmospheric and
Environmental Chemistry Harvard University MBI
workshop, Columbus, OH 05 April 2006
2

• What issues are we going to address
  (implicitly)?
• Reduction of dimensions
• Making hidden processes visible through iterative
  experimental design
• Dealing with outliers how do I find them, how
  do I know if I have a stochastic anomaly or a
  threshold event?
• Merging deductive and inductive models fusion
  of data and models.

3

• What are we trying to do ?
• Understand the emergent properties of ecosystems,
  because these are the key to assessing the
  impacts of climate change and to making wise
  management decisions. For example
• Evaluate the effects of legacies of prior land
  use and environmental change on the forests of
  North America.
• Obtain the carbon budget for North America by
  driving a data fusion product with observed
  temperatures (assimilated meteorology), sunlight
  (GOES), and MODIS data.
• Predict vegetation change and carbon cycle
  changes and feedbacks in response to climate
  change and land management.
• Who we are

Daniel M. Matross, Pathmathevan Mahadevan,
Christoph Gerbig, John C. Lin, Arlyn Andrews,
John T. Lee , J. William Munger, V. Y. Chow,
David Hollinger, Bruce Daube, Alfram Bright,
Elaine W. Gottlieb, Shawn Urbanski, Carol
Barford , Paul Moorcroft, David Medvigy, Marcos
Albani, Steven C. Wofsy
Harvard University USFS U. of Maine GMD,
NOAA U. Wisc.
4
What are emergent properties of ecosystems?
An emergent property of a system is a functional
relationship that arises from complex
interactions at a lower level of organization.
Emergent properties are created by small biases
at the individual (agent) level and involve
probability distribution functions. Example the
equation of state for a gas (PVNkT) arises from
the laws of Newtonian physics applied to
molecules moving through space it can be derived
from a probability equation describing molecular
motions (Boltzmanns Equation). Example the
temperature dependence of ecosystem respiration
arises from the temperature dependence of an
ensemble of enzymatic processes it cannot
(presently) be predicted by knowing any subset of
the underlying functional relationships. Can we
obtain the equation of state for a tree? For an
ecosystem? What are the state variables for these
systems? Note the analogy to the upper level of
an hierarchical model.
5
Example 1 Do (Why do) regenerating forests in
the US remove CO2 from the atmosphere?
Harvard Forest, central Massachusetts A
"typical New England forest" but it was not
always so! (in fact, it was never so)
6
Our legacy land use change in New England
Fitzjarrald et al., 2001
7
5x109 (x7) data points
A
Net Exchange (mmol CO2/m2/s)
-30 -20 -10 0
B
Hourly data for Net Ecosystem Exchange, 11 yrs
(A) 2 days. (B). Two days of hourly data. C. 13
years of respiration (R), GEE, and D. 13 years of
NEENEE annual sums. If data started/ended in
1998?
uptake
emission
C
D
8
Can you/Who would believe a sum of 150,000 half
hours (10 Hz!), with 25 rejected/missing data?
Design an independent test
9
Long-term C uptake 1998 anomaly
Live Biomass/ Increment yr-1
uptake
LUE1200-1500
emission
1998
10
Did anything significant happen in January, 1998,
about which we have/can get information?
11
(No Transcript)
12
The 15-year trend represents the net growth of
the forest after 65 years of succession. It is
taking up much more CO2 during this middle stage
of succession than models of forest growth had
predicted. To evaluate the trends from data, it
is critically important to identify by
independent means major perturbing factors, e.g.
the ice storm that interrupted the smooth
progression of forest development in a large
region of North America.
13

• Example 2 How can we develop the most
  reduced-dimension bottomup model, spatially and
  temporally resolved, for regional surface fluxes?
• We will use this model as the prior for a
  Bayesian inversion of the atmospheric data over
  land, to obtain regional/continental scale fluxes
  and budgets for CO2.
• Ingredients
• CO2 fluxes from eddy covariance towers across the
  USroughly 50 with good data quality and several
  years of data (hourly).
• Atmospheric concentration data from tall towers
  at sites across the continent (hourly). The
  sites are in the midst of the active surface and
  respond to both near-field and far-field
  influences.
• Atmospheric wind and temperature data (hourly)
  assimilated meteorological products.
• Hourly sunshine from the NLDAS (hourly, satellite
  data plus).
• Land cover maps at 1 km.
• Satellite reflectance data (EVI, ) at 1 km, 8
  day intervals.
• Fossil fuel and wildfire sources, from
  inventories.
• Incommensurate data all bearing on the same
  process
• a modeldata fusion problem.

14

• VPRM Vegetation Photosynthesis and Respiration
  Model
• Simple mathematical structure
• Minimum parameters 4 (?,?, PAR0, b) x 11
  vegetation classes invariant in time or space
  fit and validated using Ameriflux and Fluxnet
  Canada eddy covariance data
• Rich temporal and spatial complexity of carbon
  flux captured with remote sensing data
• Ingests vast data amts, very simple hourly
  output for users

Assimilated Meteorological Data (GOES sun)

• Tower Fluxes

• Meterological Data ltlt

Tower Data
MODIS Reflectances

• MODIS Data

LSWI
EVI
Validation
GEE ? ? (Tscalar ? Wscalar ? Phscalar )?FAPAR ?1
/(1 PAR/PAR0) x PAR
PAR? SWR Soil H2O TBA (NLDAS)
R ? ? Ta ß
Ta ? TRm
Pathmathevan Mahadevan et al., 2006 submitted to
Global Biogeochemical Cycles.
15
Existing models on a complexity scale
From Siqueira et al. (2006)
Hourly
(From G. Katuls presentation)
16
Scalars account for variations in primary
production using known functional forms
Temperature (parabolic)
Water Stress (linear)
Phenology (switch, damped)
17
Satellite-derived vegetation in NE North America,
subdivided by climate-derived life zone.
18
Calibrate this model using eddy flux data, then
test it using independent data, over and over,
until you cant kill it any more
Monthly mean NEE (mmole m-2s-1)
BLUE MODEL BLACK OBSERVATIONS Fitted to
hourly data using non-linear least-squares
19
VPRM CROSS-CALIBRATION
For Harvard
For Howland
BLUE MODEL BLACK OBSERVATIONS
20
Pathmathevan Mahadevan produced these plots
CROSS-CALIBRATION (each veg. type)
For Harvard
For Howland
BLUE MODEL BLACK OBSERVATIONS
21
What you get
VPRM a priori carbon budget for 41 to 52 N, 65
to 80 W Tg C
plus direct estimate of the error covariance
matrix, by veg type with spatial decorrelation
scale.
Monthly average NEE (µmol C/m2/s)
22
August 2004 Net CO2 Exchange mmoles C m-2 s-1
29 June 20031800 GMT mmoles C m-2 s-1
VPRM Continental Carbon Budgets Assimilated of
AmeriFlux tower data, GOES sunshine, and MODIS
EVI and LSWI The prior for a fully constrained C
budget
23
Concentrations measured at tall towers sample
both the mean field, influenced by forests many
100s of km distant, and by the near field
CMDL (Argyle, ME) tall tower
24
Model-Data FUSION VPRM Lagrangian ensemble
transport model The prior fluxes from VPRM
provide extremely accurate predictions of
large-scale CO2 concentrations for N. America.
MERGES AmeriFluxMet.MODIS
Tall tower data from Maine
Surface flux model is not adjusted to match
concentration data it is the prior model for a
Bayesian inversion to obtain large-scale
ecosystem C budgets
log10 ppm/(mmole/m2/s) 1/4 x 1/6 grid sq.
25
(No Transcript)
26
Se Sveg_mdlSparticleSeddyStransportSaggrSoc
eanScombust
Spart is the random error due to particle
statistics, Seddy is the error due to unresolved
eddies, Stransp describes the error in the
transport model, including mixed layer height and
errors in the wind fields, Saggr represents the
error due to aggregation of fluxes into gridded
spatial regions (1/4 x 1/6 deg), plus
representation error for the measurements,
Socean accounts for neglecting oceanic fluxes
Socean accounts for errors in fossil and
vegetation combustion fluxes
27
Add meas. err.
simulation
Steps required to obtain the representation
error part of the posterior error covariance
matrix.
28

• Summary and Conclusions
• The problem of understanding ecosystem function
  and dynamics in the context of climate change is
  a problem of linked scales over many orders of
  magnitude, both in the temporal domain and the
  spatial domain.
• Ecosystems have emergent properties on
  successively larger scales, in both space and
  time.
• Measuring and modeling these properties involves
  close interactions between observations and
  stochastic and mechanistic models. The analysis
  and synthesis should lead to stronger, more
  rigorous experimental designs that simultaneously
  remove arbitrary components from the models and
  reduce uncertainties in a verifiable way.
• We can use some help. To escape the false
  dichotomy of statistical vs. deterministic
  models (Clark), we need to train all our
  students (and ourselves?) more broadly to develop
  better experimental designs and synthesis
  procedures.

29
Emergent properties of ecosystems gaining
scientific knowledge from data and models at
multiple temporal and spatial scales by data
fusion that combines data from eddy fluxes, tall
towers, aircraft and satellites
CMDL (Argyle, ME) tall tower
End of presentation
30
Level of OrganizationgtgtTime and Space
Scales Time minute-to-minute gtgt day gtgt week
gtgt month gtgt year gtgt decade Process
physiology light rx transport
growth succession, climate


disturbance System organ, ensemble
plantsenviron. ecosystem ecosys.
emergent
environment
properties Space leaf gtgt tree gtgt
canopy gtgt whole ecosystem gtgt landscape gtgt
region