Progress with land DA

P. Lewis

UCL Geography & NERC NCEO

Integration of RS products into process models

Testing: evaluate model performance via diagnostic variables

Forcing: using RS estimates as new updates of state variables

Assimilation: adjusting model parameters or initial conditions so that diagnostic variables simulated by the model is close to RS estimates

What are the requirements of the models/EO?

Testing: compare ‘diagnostics’ (e.g. fAPAR, LAI)

Brut et al. 2009 BiogeosciencesISBA

MODIS

CYCLOPES

LAI as diagnostic variable for comparison

Pragmatic: rescale and smooth (‘bias’)Essentially use EO phenology

Barriers to effective DA

• Why are products different?

• Different assumptions/treatments/(datasets)

• In any case inconsistent with model assumptions …

• Radiance DA

• DA/comparisons with low level EO

• Advantages:

• ‘control’ over data interpretation assumptions

• i.e. definition of observation operator: consistency ?

• (potentially) include multiple EO (in consistent manner)

• More easily treat uncertainties++

Initial efforts: Quaife et al. 2008 (+)

T. Quaife, P. Lewis, M. DE Kauwe, M. Williams, B. Law, M. Disney, P. Bowyer (2008), Assimilating Canopy Reflectance data into an Ecosystem Model with an Ensemble Kalman Filter, Remote Sensing of Environment, 112(4),1347-1364.

Shaded crownIlluminated crown

Illuminated soil

Shaded soil

Flux tower site 1: Oregon (‘Young’)

Issues

• NEP good, but actually over-estimate GPP …

• Partial structural consistency

• ‘lumped’ ecosystem model (only one canopy layer)

• ‘effective’ LAI

• Even though tried to account for canopy-scale clumping

• Fixed parameters (e.g. leaf chlorophyll):

• Because of radiometric trade-offs between LAI & e.g. chlorophyll

• Because assume (e.g.) SLA known (& fixed)

• Because of partial structural consistency

Leaf Area Index

• Area based measure of leaf amount

• Related to mass-based through SLA• Generally assumed constant

Kattge et al., 2011

SLA

Data Assimilation and EOLDAS

• Make all terms EO sensitive to dynamic

• Weak constraint 4DVAR

Lewis et al. 2012 RSE

EOLDAS

EOLDAS++

• Follow-on project (2012-2014)

• Integrate with model (with Reading)

• Deal with snow/soil water

• Build in JULES-like vegetation model and Observation operators• Include passive microwave obs. op.

Regularisation for albedo

Quaife and Lewis 2010

2005 8-dailies BRDF f0 (SW, NIR, VIS= r,g,b)

Change detection: disturbance

Relax constraint at discontinuities

Disturbance

Disturbance

Disturbance

• Current:

• Edge-preserving DA in time

• Next

• Extend spatially

• Multiple constraints (FRE)

• Build in model interpretation• Initially FCC

• Most of tools in place to track state within DA system• To estimate biomass loss and recovery

Guanter et al. (2012) RSE accepted: Fs GOSAT

Fs Modelling

First model

• Initial exploration with regulariser

• Compare to environmental constraints

• Multi-model DA

• MPI-GPP, PEM, Fs data (linear GPP model)

• Use to constrain extrapolation of MPI observations

• Data quite noisy

• How far to go with GOSAT?

Summary

• weak constraint: regularisation (xval):

• Enable to treat ~all terms EO sensitive to• E.g. chlorophyll etc.

• Build in disturbance/change• Time/(space)

• Multiple model/data constraints • Working out how to deal with new observations (Fs)

Where next?

• EO-LDAS++

• Exploit EOLDAS ideas in model-data integration

• More observation operators & underlying process model

• Structural consistencies / learn from TRY (Terrabites)

• Disturbance DA

• Build up: spatial; FRE; FCC; process model …

• testbed & high res tracking system for C emissions and interaction with vegetation (e.g REDD+ work with Edinburgh)

• Fs?

• More testing, examining at higher spatial & time res. (DA)

• Integrate with rest of DA work

• Interface to atmospheric models?

Thank you

Conclusion

• Integration of RS products into process models

• Testing; forcing; assimilation

• Main EO role so far constraining LC & timing (phenology, snow)

• Barriers to progress …

• Model/interpretation inconsistencies / fixed parameters

• Need to work on this in observations & models

• Important tools

• Weak constraint DA

• ‘low level’ DA

Conclusions 1/2

• LAI products still not optimally used• Lack of uncertainty information

Conclusions 2/2

• Clumping major issue • Potentially accessible from EO

• Or can model impacts even in simple models

• Minimum requirement: LAI and crown cover…

• BUT do we need to deal with it? • Or is effective LAI (i.e. including clumping) sufficient?

• If so, significant implications for EO efforts• And model testing

• (Keep in mind need for direct/diffuse on fAPAR/albedo)

Thank you

Canopy Scaling of leaf process

Sellers 1992 canopy process scaling model:

• Assume leaf N, Vmax, Vm profiles distributed according to fAPAR profile

• obtain scalar from (top) leaf to canopy scale process (assim., resp, transp.)

v=0.2

Canopy Photosynthesis

If horizontal heterogeneity in LAI (Sellers, 1992)

•Spatial variation in fAPAR

•But fAPAR scales ~linearly

•If Ac,k etc constant

• Aci ~ fAPARi/k

• So Ac (etc) scale linearly

• (in the absence of variations in forcings and process rates)

•BUT does not nec. follow if other leaf process to canopy scalings assumed

• E.g. per layer A in JULES

• Different assumptions about leaf N vertical distribution

scattering asymmetry: impact

Often assume scattering isotropic

For diffuse fluxes, e.g. -Eddington

radiatively account for

asymmetry in phase fn

by mapping to equivalent LAI and

e.g. NIR: =0.9, e.g. f=0.3, L’=0.73L, ’=0.86

f: fractional scattering in fwd peak

Problems in the retrieval of variables : balance between accuracy and precision

32

Best reflectance match Median of cases within ±1σ

Very little bias but large scattering

Accurate, not precise

Smaller scattering but larger biases

Precise, not (always) accurate

Selection of solutions within a Look Up Table (LUT). Measurement uncertainties

Importance of: - the retrieval method

- knowledge of uncertainties (model and measurements)

- the prior distribution of input variables

Importance of: - the retrieval method

- knowledge of uncertainties (model and measurements)

- the prior distribution of input variables

LAI anomalies

Shoot-scale clumping reduces apparent LAI

Smolander & Stenberg RSE 2005

pshoot=0.47 (scots pine)

p2<pcanopy

Shoot-scale clumping reduces apparent LAI

Scaling properties

Weiss et al. 2000

Differences depending on directionality

Clumping impact

Govind et al. 2010

Also interest in non-photosynthetic vegetation e.g. for fire