Transcript Slide 1

EO data assimilation in land process
models
Mat Disney and Shaun Quegan
No one trusts a model except the man who wrote it; everyone trusts
an observation except the man who made it (Harlow Shapley)
Concept for Global Carbon Data Assimilation System
NB carbon and water are inextricably linked, so this is a more
generalised vegetation – soil – water- atmosphere scheme
Geo-referenced
Geo-referenced
emissions
emissionsinventories
inventories
Climate and weather
fields
Ocean time series
Biogeochemical
pCO2
Surface
observation
pCO2
nutrients
Water column
inventories
Atmospheric
Atmospheric
measurements
measurements
Remote
Remote sensing
sensing of
of
atmospheric
atmospheric CO
CO22
Atmospheric
Atmospheric Transport
Transport
Model
Model
Ocean
Ocean Carbon
Carbon
Model
Model
Coastal
Coastal
studies
studies
Optimised
Optimised
fluxes
fluxes
Terrestrial
Terrestrial
Carbon
Carbon Model
Model
rivers
Lateral fluxes
Data
assimilation
link
Optimised
Optimised
model
model
parameters
parameters
Eddy-covariance
flux towers
Biomass soil
carbon
inventories
Ecological
studies
Ocean remote sensing
Ocean colour
Altimetry
Winds
SST
SSS
Remote sensing of
vegetation properties
Growth cycle
Fires
Biomass
Radiation
Land cover/use
Ciais et al. 2003 IGOS-P Integrated Global Carbon Observing Strategy
Terrestrial Component
Remote
Remote sensing
sensing of
of
atmospheric
CO
atmospheric CO22
Climate and weather
fields
Atmospheric
Atmospheric Transport
Transport
Model
Model
Optimised
Optimised
fluxes
fluxes
Terrestrial
Terrestrial
Carbon
Carbon Model
Model
rivers
Which model(s)
should go here?
Lateral fluxes
Optimised
Optimised
model
model
parameters
parameters
Eddy-covariance
flux towers
Biomass soil
carbon
inventories
Ecological
studies
+ Water components:
SWE
soil moisture
Remote sensing of
vegetation properties
Growth cycle
Fires
Biomass
Radiation
Land cover/use
Land process models
Land models need to deal with transfers of
- energy
- matter
- momentum
between the land surface and the atmosphere.
Three classes of land (coupled carbon-water) models:
 Models driven by radiation (light use efficiency models)
 Dynamic Vegetation Models: climate driven
 Simple box models
Some models emphasise hydrology (not discussed here)
Light Use Efficiency models
Incoming
PAR
CO2
Absorption
fAPAR
LUE
Photosynthesis
Respiration
GPP
NPP
Efficiency coefficient: LUE
365
NPP  LUE .  fAPAR . PAR
day 1
The LUE may depend on biome, soil
moisture, temperature, nutrients, age,
Measured by
satellites
Modeled or
measured by
satellites
Notes on LUE models
 Models built by ecologists tend to focus on
leaves as the functional element (e.g. Leaf
Area Index).
 Models built by remote sensors tend to focus
on radiation.
 LUE models are driven by EO data, rather than
geared to assimilating data.
Properties of DVMs
DVMs originally designed to examine long-term
trends under climate change so…
 Data-independent, except for varying climate
data and static soil texture data
 Comprehensive description of biophysics
 All processes internalised, parameterised
 Complex, non-linear, non-differentiable,
(discontinuities, thresholds)
 Expensive to run
The Structure of a Dynamic Vegetation Model
Parameters
Climate
Sn
Soil
texture
DVM
Processes
Sn+1
Testing
How EO data can affect DVM calculations
Phenology
Snow water
Burnt area
fAPAR
Parameters
Processes
Climate
Land
cover
Forest age
Sn
Soils
DVM
Sn+1
Observable
Possible feedback
Testing:
Radiance
fAPAR
Calibration– boreal budburst
Offline setting of global parameters can be thought
of as a form of DA, but is better described as
model calibration.
In the following e.g, we use new EO observations
that are unaffected by snow-melt to
parameterise the spring warming boreal
phenology model.
The SDGVM budburst algorithm
When
 min(0, T – T0) > Threshold, budburst occurs.
days
The sum is the red area. Optimise over the 2 parameters, Threshold
and T0 (minimum effective temperature).
T0
Start of budburst
The Date of budburst derived from
minimum NDWI (VGT sensor, 2000) N. Delbart, CESBIO
Day of year
Testing SDGVM with EO data
SDGVM can predict satellite ‘observations’ since it
contains a canopy model and the concept of
radiation interception
Model “skill”
1999
SDGVM fAPAR
AVHRR NDVI
Bad
Skill
Good
Are derived parameters the problem?
Is the problem the SDGVM or the derived
parameter from the EO signal?
The next slide shows the fAPAR derived from
Seawifs (JRC) and from MODIS for a site in the
UK. The large bias between the two is a
general feature of these two datasets.
Biases in derived parameters
Assimilating products
Assumptions
Observations
Assumptions
Observations
Data Assimilation Scheme
(KF, EnKF, 4DVAR, etc)
MODEL
Assumptions
For example: soil moisture
from SMOS, surface
temperature, LAI from MODIS
Low-level vs derived products
 similar products give substantially different
values;
 assumptions used to derive products usually
inconsistent with biospheric models;
 Product uncertainties are poorly known
 Can we use low-level products (Reflectance?
BOA radiance? TOA radiance?)
Assimilating reflectance
Data Assimilation Scheme
(KF, EnKF, 4DVAR, etc)
Observations
Observations
Observation Operator
Assumptions
e.g. reflectance,
backscatter, etc…
MODEL
Assumptions
Assumptions in the observation
operator are made to be consistent
with those in the model
Observation operators
This approach needs observation operators: translate
ecosystem model state vector into observable
properties e.g.
 reflectance data assimilated into DALEC;
 predicting radar coherence in ERS Tandem data from
the SPA model;
 relating snowpack properties to SSM/I radiometer data;
 recognising burnt area and severity of burn.
Which is the right model?
 Complex DVM-type models never designed for
DA
 So, pursuing another approach with a simplified
box model designed from the start for DA
– DALEC
The Structure of a Data Assimilation Model (DALEC)
EO data
(e.g. LAI, VI,
reflectance)
Blue lines indicate
integration of EO
data with DALEC
Ensemble
Kalman
Filter
Observation
model
Ra
Ppt
Rh
ET
WS1
Cf
Cl
GPP
Cr
Cw
WS2
Cs
Q
WSn
Stocks and fluxes of carbon (left) and water (right)
Observation operator: simple RT model + snow
Canopy foliage results
No assimilation
Assimilating
MODIS
(bands 1 and 2)
Canopy foliage results
Assimilating
MODIS
exc. snow
Assimilating
MODIS
inc. snow
Quaife, Williams, Disney et al. RSE in press
EO land cover and carbon
 All EO land cover the same?
 DGVMs use land cover indirectly
– How do we translate land cover classes to
PFTs?
Quaife, Quegan,
Disney et al.,
submitted
EO land cover and carbon
Quaife, Quegan, Disney et al., submitted
How do we find best model-data framework?
 Use ‘God’ models to test assumptions of simpler
models
– DVMs + DALEC-type models
 Model-data fusion inter-comparison e.g. REFLEX:
Regional Flux Estimation Experiment
– www.carbonfusion.org
– Compare strengths/weaknesses of various modeldata fusion techniques
– Quantify errors/biases introduced when
extrapolating fluxes in both space and time using a
model constrained by model-data fusion methods.
Key issues for DA in land models 1
 Models
– Simple enough for effective DA but complex
enough to capture biophysics
– Suitable interface with observation
operators
– preferably differentiable
Key issues for DA in land models 2
 Data
– Same meaning of observed parameters as used in
models
– Proper characterisation of uncertainty i.e. PDFs
– Use OOs to make best use of all available data e.g.
optical, LiDAR, RADAR, thermal ….
 We are still searching for the best model-data structure.
Key issues for DA in land models 3
 DA through observation operators not only
answer, for various practical reasons.
 Also pursue general concepts of how EO data
can reduce the uncertainty in land models
– Calibration, testing etc.
Thank you
Severity of disagreement – AVHRR/SDGVM
1998
r > 0.497 OR r.m.s.e < 0.2
r < 0.497 AND r.m.s.e > 0.2
r < 0.497 AND r.m.s.e > 0.3
Severity of disagreement – example
Mid Europe
Severity of disagreement – example
SW China
Lesson
1. The DVM as currently formulated only supports a
simple observation operator. This allows meaningful
estimates of time series of observables; absolute
values of the observables are of dubious value.
2. These time series permit the model to be interrogated
with satellite data, and model failures to be identified.
Detecting incorrect land cover
Crop class incorrectly set
Crop class correctly set
0.9
0.0
Pearson’s product moment
Temporal correlation
Lesson
Forward operators may prove a powerful tool in land cover
mapping
Impact on Carbon Calculations
Calibrated model is unbiased,
unlike methods based on NDVI
1 day advance: NPP increases by 10.1 gCm-2yr-1
15 days advance: 38% bias in annual NPP
Observations
calibrate
Carbon Calculation
Phenology model
Picard et al.,GCB, 2005
Dynamic Vegetation
Model
Comparison Model-EO: RMSE
Model needs to be region specific,
here include chilling requirement ?
NDVI predicted by SDGVM
1998
0
1999
1
0
1
A Dynamic Vegetation Model (SDGVM)
ATMOSPHERIC
CO2
Photosynthesis
Fire
GPP
GROWTH
BIOPHYSICS
Mortality
NPP
Thinning
Litter
Soil
LEACHED
NBP
Biomass
Disturbance
Assimilating reflectance
Observations
Data Assimilation Scheme
(KF, EnKF, 4DVAR, etc)
MODEL
The real world
Assumptions
But how do we use a nonlinear observation operator?
Comparing model and measured fAPAR
August 99
May 99
Seawifs
SDGVM
Model and predicted fAPAR
Average over
the whole of
Europe for 1999
and 2000
Note: if SDGVM
were driven by
the Seawifs
values, most
model forests
would die
Experiments
 State and parameter estimation. DE1 and EV1 sites, 3
years driving data, all available obs
 As 1. but using synthetic data (DE2 and EV2)
 Within site forecasting. Another year of driving data for
DE1 and EV1, but no observations
 As 3. but using synthetic data (DE2 and EV2)
 Between site extrapolation. DE3 and EV3 sites, 4 years
driving data, MODIS LAI only
Integrated flux predictions
Flux
(gC.m-2)
NEP
GPP
Assimilated
data
Total
Standard
Deviation
Assimilation
exc. snow
373.0
151.3
Assimilation
inc. snow
404.8
129.6
Williams et al.
(2005)
406.0
27.8
Assimilation
exc. snow
2620.3
96.8
Assimilation
inc. snow
2525.6
42.7
Williams et al.
(2005)
2170.3
18.1
REFLEX data sets
 “Paired” sites to test extrapolation/estimation
– Brasschaat (DE2) and Vielsalm (EV2) (MF)
– Hainich (DE3) and Hesse (DE1) (DBF)
– Loobos (EV1) and Tharandt (EV3) (ENF)
 Meteorological drivers, fluxes, MODIS LAI and
stocks
– Attempting to estimate “uncertainty” in
fluxes and MODIS LAI
REgional Flux Estimation eXperiment (REFLEX)
FluxNet data
MODIS
Training Runs
Assimilation
MDF
DALEC
model
Output
Full analysis
Model parameters
Deciduous forest sites
Coniferous forest sites
REgional Flux Estimation eXperiment (REFLEX)
FluxNet data
MODIS
MDF
Testing site forecasts
with limited EO data
DALEC
model
FluxNet data
testing
Full analysis
Model parameters
MDF
Assimilation
MODIS
Analysis