Dr. Yiqi Luo Botany and microbiology department University of Oklahoma, USA

34
Parameter identifiability, Parameter identifiability, constraints, and constraints, and equifinality in data equifinality in data assimilation with ecosystem assimilation with ecosystem models models Dr. Yiqi Luo Dr. Yiqi Luo Botany and microbiology Botany and microbiology department department University of Oklahoma, University of Oklahoma, USA USA Land surface models and FluxNET data Edinburgh, 4-6 June 2008 (Luo et al. Ecol Appl. (Luo et al. Ecol Appl. In press In press ) )

description

Parameter identifiability, constraints, and equifinality in data assimilation with ecosystem models. (Luo et al. Ecol Appl. In press ). Dr. Yiqi Luo Botany and microbiology department University of Oklahoma, USA. Land surface models and FluxNET data Edinburgh , 4-6 June 2008. - PowerPoint PPT Presentation

Transcript of Dr. Yiqi Luo Botany and microbiology department University of Oklahoma, USA

Page 1: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

Parameter identifiability, Parameter identifiability, constraints, and equifinality constraints, and equifinality

in data assimilation with in data assimilation with ecosystem modelsecosystem models

Dr. Yiqi LuoDr. Yiqi LuoBotany and microbiology Botany and microbiology

departmentdepartmentUniversity of Oklahoma, USAUniversity of Oklahoma, USA

Land surface models and FluxNET dataEdinburgh, 4-6 June 2008

(Luo et al. Ecol Appl. (Luo et al. Ecol Appl. In pressIn press))

Page 2: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6

x 10-4

0

50

100

150

200

250

300

350

400

Histogram of generated samples for c2

Range of c2

Sam

plin

g fre

quen

cy

0.006 0.008 0.01 0.012 0.014 0.016 0.018 0.02 0.022 0.024 0.0260

100

200

300

400

500

600

Histogram of generated samples for c3

Range of c3

Sam

plin

g fre

quen

cy

Observed Data

Prior knowledge Posterior distribution

3 3.5 4 4.5 5 5.5 6 6.5

x 10-3

0

100

200

300

400

500

600

700

800

Histogram of generated samples for c5

Range of c5

Sam

plin

g fre

quen

cy

Parameter identifiability

Inverse model

Constrained

Edge-hitting

Equifinality

Page 3: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

Wang et al. (2001) ------ a maximum of Wang et al. (2001) ------ a maximum of 3 or 43 or 4 p parameters can be determined.arameters can be determined. Braswell et al. (2005) ------ Braswell et al. (2005) ------ 13 out of 2313 out of 23 parame parameters were well-constrained.ters were well-constrained. Xu et al. (2006) ------ Xu et al. (2006) ------ 4 or 3 out of 74 or 3 out of 7 parameters parameters can be constrained, respectively under ambiecan be constrained, respectively under ambient and elevated COnt and elevated CO22..

Identiable parameters

Page 4: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

Three methods to exaThree methods to examine parameter identimine parameter identifiabilityfiability1.1. Search methodSearch method2.2. Model Model

structurestructure3.3. Data variabilityData variability

Page 5: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

Harvard Forest EMS-Tower

Eddy flux data Eddy flux data

Page 6: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

COCO22 flux flux HH22O fluxO flux Wind speedWind speed TemperatureTemperature PARPAR Relative humidityRelative humidityHourly or half-hourlyHourly or half-hourly

Eddy flux technology

Page 7: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

Leaf-level Photosynthesis

Sub-model

Canopy-level Photosynthesis

Sub-model

System-level C balanceSub-model

ModelModel

Page 8: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

Table 1 Parameters informationTable 1 Parameters information

Page 9: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

Develop prior distributionDevelop prior distribution

Apply Metropolis-Hasting algorithmApply Metropolis-Hasting algorithm

a) generate candidate a) generate candidate pp from sample space from sample spaceb) input to model and calculate cost functionb) input to model and calculate cost functionc) select according to decision criterionc) select according to decision criteriond) repeatd) repeat

Construct posterior distributionConstruct posterior distribution

Bayesian inversionBayesian inversion

Page 10: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA
Page 11: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA
Page 12: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

Conditional Bayesian inversionConditional Bayesian inversion

Bayesian inversion

Bayesian inversion

Bayesian inversion

Bayesian inversion

Page 13: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA
Page 14: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

Fig. 2 Decrease of cost function with each step of conditional inversion

Page 15: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA
Page 16: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

ConclusionsConclusions Conditional inversion can Conditional inversion can

substantially increase the number of substantially increase the number of constrained parameters.constrained parameters.

Cost function and information loss Cost function and information loss decrease with each step of conditional decrease with each step of conditional inversion.inversion.

Page 17: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

Measurement errors Measurement errors and parameter identiand parameter identifiabilityfiability

Page 18: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

Leaves X1 Woody X2 Fine Roots X3

Metabolic Litter X4 Structural Litter X5

Microbes X6

Slow SOM X7

Passive SOM X8

GPP

TECO – biogeochemical model

Page 19: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

)(

100000010000

1000000101000010000001000000001000000001

8786

7675

68676564

5351

4341

cdiagC

ffff

ffffffff

A

0)0(

)()()(

XtX

tBPtACXtXdtd

TbbbB )00000( 321

No. of parameter

8

12

8

3

Page 20: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

c1

010

2030

40

0.0E+00 1.0E-03 2.0E-03 3.0E-03 4.0E-03 5.0E-03

2.0-SD

1.0-SD

0.5-SD

c2

0

5

10

15

0 0.00005 0.0001 0.00015 0.0002 0.00025

2.0-SD

1.0-SD

0.5-SD

c3

05

10152025

0 0.002 0.004 0.006 0.008

2.0-SD

1.0-SD

0.5-SD

c4

0

1

2

3

4

0 0.01 0.02 0.03 0.04

c5

0

2

4

6

0 0.001 0.002 0.003

c6

0

20

40

60

80

0 0.1 0.2 0.3 0.4 0.5

c7

0

2

4

6

8

10

12

0 0.0005 0.001 0.0015

c8

0

0.5

1

1.5

2

2.5

0 2E-06 4E-06 6E-06 8E-06 0.00001

Exit rates

Page 21: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

A1

0

1

2

3

0 0.5 1 1.5

2.0-SD

1.0-SD

0.5-SD

A4

012345

0 0.2 0.4 0.6 0.8 1

2.0-SD

1.0-SD

0.5-SD

A5

0

1

2

3

4

0 0.2 0.4 0.6

2.0-SD

1.0-SD

0.5-SD

A6

0

1

2

3

0 0.2 0.4 0.6

A7

0

1

2

3

0 0.2 0.4 0.6 0.8

A8

0

1

2

3

4

0 0.1 0.2 0.3 0.4

A9

0

1

2

3

0 0.2 0.4 0.6 0.8

A10

0

1

2

3

0 0.1 0.2 0.3 0.4

A11

0

1

2

3

0 0.5 1 1.5

Transfer coefficients

Page 22: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

Xo1

02

46

8

0 100 200 300 400 500

2.0-SD

1.0-SD

0.5-SD

Xo2

0

510

15

20

0 2000 4000 6000

2.0-SD

1.0-SD

0.5-SD

Xo3

0

12

3

4

0 100 200 300 400

2.0-SD

1.0-SD

0.5-SD

Xo4

00.5

11.5

22.5

0 20 40 60 80

Xo5

0

1

2

3

4

0 100 200 300 400 500

Xo6

0

1

2

3

0 50 100 150

Xo7

0

5

10

15

20

0 1000 2000 3000 4000

Xo8

0

5

10

15

20

0 200 400 600 800

Initial values

Page 23: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

X1

500 1000 1500 2000

Freq

uenc

y 10

2

05

10152025

X2

5000 6000 7000 8000 900001234

X3

500 1000 1500 2000

Freq

uenc

y 10

2

05

10152025 X4

500 1000 1500 20000

10203040

X5

500 1000 1500 2000

Freq

uenc

y 10

2

02468

10 X6

50 100 150 2000

20

40

X7

Carbon content (g C m-2)1000 2000 3000 4000

Freq

uenc

y 10

2

02468

10X8

Carbon content (g C m-2)

450 500 550 600 65002468

Pool sizes without data

Page 24: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

X1

300 400 500 600 700 800

Freq

uenc

y 10

2

05

10152025

X2

5000 6000 7000 8000 900005

10152025

X3

100 200 300 400 500

Freq

uenc

y 10

2

05

10152025

X4

100 200 300 400 50005

10152025

X5

0 500 1000 1500 2000

Freq

uenc

y 10

2

05

10152025 X6

0 50 100 150 20005

10152025

X7

Carbon content (g C m-2)1500 2000 2500 3000 3500

Freq

uenc

y 10

2

05

10152025

halved SDambient SDdoubled SD

X8

Carbon content (g C m-2)

700 800 900 1000 1100 1200 130005

10152025

Pool sizes with data and different SD

Page 25: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

ConclusionConclusion

Magnitudes of measurement errMagnitudes of measurement errors do not affect parameter idenors do not affect parameter identifiability but influence relative tifiability but influence relative constraints of parametersconstraints of parameters

Page 26: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

Base modelBase modelGPP

Leaves X1 Stems X2 Roots X3

Metabolic L. X4 Struct. L. X5

Microbes X6

Slow SOM X7

Passive SOM X8

Page 27: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

Simplified modelsSimplified models

Plant C

Litter C

GPP CO2

Soil C

Plant C

Litter C

GPP CO2

O Soil C

Miner. C

3P model 4P model

Page 28: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

Simplified modelsSimplified models6P model 7P model

GPP

Leaves X1 Stems X2 Roots X3

Litter X4

Slow C X5

Miner. Soil C X6

GPP

Leaves X1 Stems X2 Roots X3

Metabolic L. X4

Struct. L. X5

Microbes X6

Soil C X7

Page 29: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

3P model-parameter 3P model-parameter constraintsconstraints

c1

0

100

200

300

400

500

600

2.08

E-04

2.27

E-04

2.46

E-04

2.65

E-04

2.83

E-04

3.02

E-0

4

3.21

E-04

3.40

E-04

3.58

E-0

4

3.77

E-04

3.96

E-04

4.15

E-0

4

4.33

E-04

4.52

E-04

4.71

E-0

4

c2

0100200300400500600700

1.56

E-03

1.89

E-03

2.22

E-0

3

2.55

E-03

2.88

E-03

3.21

E-03

3.54

E-03

3.87

E-03

4.20

E-03

4.53

E-03

4.86

E-03

5.19

E-03

5.53

E-03

5.86

E-03

6.19

E-03

c3

0

100

200

300

400

500

600

4.80

E-0

5

6.79

E-05

8.78

E-05

1.08

E-04

1.28

E-04

1.48

E-04

1.67

E-0

4

1.87

E-04

2.07

E-04

2.27

E-04

2.47

E-04

2.67

E-0

4

2.87

E-04

3.07

E-04

3.27

E-0

4

Plant C Litter C

Soil C

Page 30: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

4P model-parameter 4P model-parameter constraintsconstraints

Plant C Litter C

Slow Soil C

c1

0100200300400500600700

2.81

E-0

4

3.08

E-04

3.34

E-0

4

3.61

E-04

3.88

E-0

4

4.15

E-04

4.42

E-04

4.69

E-04

4.96

E-04

5.23

E-04

c2

0200400600800

10001200

2.46

E-03

3.20

E-03

3.94

E-0

3

4.68

E-03

5.42

E-03

6.16

E-03

6.90

E-03

7.64

E-03

8.38

E-03

9.12

E-03

c3

0200400600800

1000

3.16

E-04

3.51

E-04

3.86

E-0

4

4.22

E-04

4.57

E-04

4.93

E-04

5.28

E-04

5.63

E-04

5.99

E-04

6.34

E-04

c4

0

500

1000

1500

2000

4.00

E-06

1.61

E-05

2.82

E-0

5

4.03

E-05

5.23

E-05

6.44

E-05

7.65

E-05

8.85

E-05

1.01

E-04

1.13

E-04

Passive Soil C

Page 31: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

6P model-parameter 6P model-parameter constraintsconstraints

Foliage

Litter C Slow Soil C Passive Soil C

c1

050

100

150200

1.26

E-03

1.33

E-03

1.39

E-03

1.46

E-03

1.53

E-03

1.59

E-03

1.66

E-03

1.73

E-03

1.79

E-03

1.86

E-03

c2

050

100150200250

1.16

E-05

2.87

E-0

5

4.58

E-05

6.30

E-05

8.01

E-05

9.72

E-05

1.14

E-04

1.31

E-04

1.49

E-0

4

1.66

E-04

c3

050

100150200250

4.37

E-03

4.54

E-03

4.70

E-03

4.87

E-03

5.04

E-03

5.21

E-03

5.37

E-03

5.54

E-03

5.71

E-03

5.88

E-03

Woody Fine roots

c4

050

100150200250300

2.52

E-03

3.00

E-03

3.49

E-03

3.98

E-0

3

4.47

E-03

4.96

E-03

5.45

E-03

5.94

E-03

6.42

E-03

6.91

E-03

c5

050

100150200250

3.61

E-04

3.95

E-0

4

4.30

E-04

4.65

E-04

5.00

E-04

5.35

E-04

5.70

E-04

6.05

E-04

6.40

E-0

4

6.74

E-04

c6

0100

200300

400

3.10

E-06

1.23

E-05

2.15

E-05

3.07

E-05

3.99

E-05

4.91

E-05

5.83

E-05

6.75

E-05

7.67

E-05

8.58

E-05

Page 32: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

7PM model-parameter 7PM model-parameter constraintsconstraintsFoliage

Metabolic L. C Structure L. C Microbes C

Woody Fine roots

Soil C

c1

050

100150200250

1.25

E-0

3

1.32

E-0

3

1.40

E-0

3

1.47

E-0

3

1.54

E-0

3

1.62

E-0

3

1.69

E-0

3

1.76

E-0

3

1.84

E-0

3

1.91

E-0

3

c2

050

100150200250300

1.16

E-0

5

2.85

E-0

5

4.53

E-05

6.22

E-0

5

7.90

E-05

9.59

E-0

5

1.13

E-04

1.30

E-04

1.47

E-0

4

1.63

E-0

4

c3

050

100150200250300

4.25

E-0

3

4.42

E-0

3

4.59

E-03

4.77

E-0

3

4.94

E-03

5.12

E-0

3

5.29

E-03

5.46

E-03

5.64

E-0

3

5.81

E-0

3

c4

0

50

100

150

3.26

E-0

3

5.97

E-0

3

8.69

E-03

1.14

E-0

2

1.41

E-02

1.68

E-0

2

1.95

E-02

2.23

E-02

2.50

E-0

2

2.77

E-0

2

c5

050

100150200250300

1.30

E-0

4

3.49

E-0

4

5.67

E-04

7.85

E-0

4

1.00

E-03

1.22

E-0

3

1.44

E-03

1.66

E-03

1.88

E-0

3

2.10

E-0

3

c6

050

100150200250300

5.24

E-03

6.19

E-03

7.14

E-03

8.08

E-03

9.03

E-03

9.98

E-03

1.09

E-02

1.19

E-02

1.28

E-02

1.38

E-02

c7

050

100150200250

8.40

E-06

2.66

E-05

4.48

E-05

6.30

E-05

8.12

E-05

9.94

E-05

1.18

E-04

1.36

E-04

1.54

E-04

1.72

E-04

Page 33: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

8P model-parameter 8P model-parameter constraintsconstraints

c1

0200400600800

9.92

E-04

1.26

E-03

1.54

E-03

1.81

E-03

2.08

E-0

3

2.35

E-03

2.62

E-03

2.90

E-03

c2

0100200300400500600

1.47

E-05

5.04

E-05

8.60

E-05

1.22

E-04

1.57

E-0

4

1.93

E-04

2.29

E-0

4

2.64

E-0

4

c3

0200400600800

3.82

E-0

3

4.29

E-03

4.75

E-0

3

5.22

E-03

5.68

E-03

6.15

E-03

6.62

E-03

7.08

E-03

c4

050

100150200250300

2.45

E-03

6.18

E-03

9.91

E-03

1.36

E-0

2

1.74

E-02

2.11

E-02

2.48

E-02

2.86

E-02

c5

0100200300400

1.47

E-04

5.05

E-04

8.62

E-04

1.22

E-03

1.58

E-0

3

1.93

E-03

2.29

E-0

3

2.65

E-0

3

c6

0500

10001500200025003000

9.36

E-03

3.52

E-02

6.10

E-02

8.67

E-02

1.13

E-01

1.38

E-01

c7

0100200300400500600

4.28

E-05

1.21

E-04

2.00

E-04

2.78

E-04

3.57

E-04

4.35

E-04

5.13

E-04

5.92

E-04

c8

0100200300400

5.00

E-07

1.70

E-06

2.90

E-06

4.10

E-06

5.30

E-06

6.50

E-06

7.60

E-06

8.80

E-06

Page 34: Dr.    Yiqi Luo Botany and microbiology department University of Oklahoma, USA

ConclusionConclusion

Differences in model structure Differences in model structure are corresponding to different are corresponding to different sets of parameters. The number sets of parameters. The number of constrained parameters varies of constrained parameters varies with data availability with data availability