bayesian inference of a dynamic vegetation model …...bayesian inference of a dynamic vegetation...
TRANSCRIPT
![Page 1: Bayesian inference of a dynamic vegetation model …...Bayesian inference of a dynamic vegetation model for grassland Julien Minet*, Université de Liège, Arlon Campus Environnement,](https://reader034.vdocuments.site/reader034/viewer/2022050413/5f89a5fb7e6dfa50393f8bca/html5/thumbnails/1.jpg)
Bayesian inference of a dynamic vegetation model for grassland
Julien Minet*, Université de Liège, Arlon Campus Environnement, BelgiumEric Laloy, SCK-CEN, Belgium
Bernard Tychon, Université de Liège, Arlon Campus Environnement, BelgiumLouis François, Université de Liège, UMCCB, Belgium
![Page 2: Bayesian inference of a dynamic vegetation model …...Bayesian inference of a dynamic vegetation model for grassland Julien Minet*, Université de Liège, Arlon Campus Environnement,](https://reader034.vdocuments.site/reader034/viewer/2022050413/5f89a5fb7e6dfa50393f8bca/html5/thumbnails/2.jpg)
MACSUR mid-term – Sassari – 1-4 April 2014
Framework ● MACSUR task L2.4 Grassland model intercomparison, G. Bellocchi.
● Phase 1 : Blind runs (end in January 2014)
● Phase 2 : Calibrated runs (on-going) : how to calibrate ?
→Calibration of a grassland model (CARAIB) by a Bayesian method
![Page 3: Bayesian inference of a dynamic vegetation model …...Bayesian inference of a dynamic vegetation model for grassland Julien Minet*, Université de Liège, Arlon Campus Environnement,](https://reader034.vdocuments.site/reader034/viewer/2022050413/5f89a5fb7e6dfa50393f8bca/html5/thumbnails/3.jpg)
MACSUR mid-term – Sassari – 1-4 April 2014
The grassland model : CARAIB
● See talk of Louis François on Wednesday 3rd April● Focused on grassland ● New management functions for grassland: cut & grazing
Reference website: http://www.umccb.ulg.ac.be/Sci/m_car_e.html
![Page 4: Bayesian inference of a dynamic vegetation model …...Bayesian inference of a dynamic vegetation model for grassland Julien Minet*, Université de Liège, Arlon Campus Environnement,](https://reader034.vdocuments.site/reader034/viewer/2022050413/5f89a5fb7e6dfa50393f8bca/html5/thumbnails/4.jpg)
MACSUR mid-term – Sassari – 1-4 April 2014
The sites
● Semi-natural grasslands : grazed (Laqueuille) or cut (other 3)
● Eddy-covariance sites : flux measurements available : GPP, RECO, ET, ...
Grillenburg, DE, 375 m
Oensingen, CH, 450 m
Monte-Bondone, IT, 1550 m
Laqueuille, FR, 1040 m
Monte Bondone (fluxnet.org)
Grillenburg (fluxnet.org)
![Page 5: Bayesian inference of a dynamic vegetation model …...Bayesian inference of a dynamic vegetation model for grassland Julien Minet*, Université de Liège, Arlon Campus Environnement,](https://reader034.vdocuments.site/reader034/viewer/2022050413/5f89a5fb7e6dfa50393f8bca/html5/thumbnails/5.jpg)
MACSUR mid-term – Sassari – 1-4 April 2014
CHAIN OF PARAMETERS
The algorithm : DREAM_ZS ● Inverse problem:
● DREAM_ZS: a Markov-Chain Monte-Carlo sampler
● Ideal for sampling a large number of parameters
● Multiple-chain : deal with local minima and correlation between parameters.
Laloy, E., and J.A. Vrugt, High-dimensional posterior exploration of hydrologic models using multiple-try DREAM_(ZS) and high-performance computing, Water Resources Research, 48, W01526, 2012
Vrugt, J.A., C.J.F. ter Braak, C.G.H. Diks, D. Higdon, B.A. Robinson, and J.M. Hyman, Accelerating Markov chain Monte Carlo simulation by differential evolution with self-adaptive randomized subspace sampling, International Journal of Nonlinear Sciences and Numerical Simulation, 10(3), 273-290, 2009.
POSTERIOR DISTRIBUTIONS
Optimal parameters=argmin (Observations−Modeled ( parameters))
![Page 6: Bayesian inference of a dynamic vegetation model …...Bayesian inference of a dynamic vegetation model for grassland Julien Minet*, Université de Liège, Arlon Campus Environnement,](https://reader034.vdocuments.site/reader034/viewer/2022050413/5f89a5fb7e6dfa50393f8bca/html5/thumbnails/6.jpg)
MACSUR mid-term – Sassari – 1-4 April 2014
The algorithm : DREAM_ZS ● Inverse problem:
● 12 parameters were sampled using 3 measurements variables from Eddy covariance: RECO, GPP, ET
● A multi-objective cost function (CF) was used : CF = f(RECO, URECO, GPP, UGPP, ET, UET)
● Uncertainties on measurement U were considered as follow (homoscedastic):
Optimal parameters=argmin (Observations−Modeled ( parameters))
Meas. variables U
RECO 1.5 gC m-2 day-1
GPP 3 gC m-2 day-1
ET 1 gC m-2 day-1
![Page 7: Bayesian inference of a dynamic vegetation model …...Bayesian inference of a dynamic vegetation model for grassland Julien Minet*, Université de Liège, Arlon Campus Environnement,](https://reader034.vdocuments.site/reader034/viewer/2022050413/5f89a5fb7e6dfa50393f8bca/html5/thumbnails/7.jpg)
MACSUR mid-term – Sassari – 1-4 April 2014
Results : parameter samplingsPosterior distributions of 9 parameters, Oensingen
![Page 8: Bayesian inference of a dynamic vegetation model …...Bayesian inference of a dynamic vegetation model for grassland Julien Minet*, Université de Liège, Arlon Campus Environnement,](https://reader034.vdocuments.site/reader034/viewer/2022050413/5f89a5fb7e6dfa50393f8bca/html5/thumbnails/8.jpg)
MACSUR mid-term – Sassari – 1-4 April 2014
Results : parameter samplings Specific leaf area (SLA) [m²/gC]
Mean = 0.0285Std = 0.0016 Mean = 0.0152
Std = 0.0005
Mean = 0.0217Std = 0.0036
Mean = 0.0149Std = 0.0009
![Page 9: Bayesian inference of a dynamic vegetation model …...Bayesian inference of a dynamic vegetation model for grassland Julien Minet*, Université de Liège, Arlon Campus Environnement,](https://reader034.vdocuments.site/reader034/viewer/2022050413/5f89a5fb7e6dfa50393f8bca/html5/thumbnails/9.jpg)
MACSUR mid-term – Sassari – 1-4 April 2014
Results : parameter samplings Specific leaf area [m²/gC]
Mean = 0.0310Mode = 0.0274Std = 0.0060
Mean = 0.0184Mode = 0.0156Std = 0.0083
Mean = 0.0279Mode = 0.0292Std = 0.0115
Mean = 0.0173Mode = 0.0141Std = 0.0090
![Page 10: Bayesian inference of a dynamic vegetation model …...Bayesian inference of a dynamic vegetation model for grassland Julien Minet*, Université de Liège, Arlon Campus Environnement,](https://reader034.vdocuments.site/reader034/viewer/2022050413/5f89a5fb7e6dfa50393f8bca/html5/thumbnails/10.jpg)
MACSUR mid-term – Sassari – 1-4 April 2014
Results : parameter samplings Specific leaf area (SLA) [m²/gC]
● SLA in CARAIB : effective SLA for a plant functional type ! ● Actually, SLA is variable between leaves and along the season ● SLA is known to depend on aridity (-) and intensification (+)
Mon
te-
Bon
done
Gril
lenb
urg
Laqu
euill
e
Oen
sing
en
De Martonne-Gottmann aridity indexfrom Ma et al. iEMSs, 2014
ARIDITY
![Page 11: Bayesian inference of a dynamic vegetation model …...Bayesian inference of a dynamic vegetation model for grassland Julien Minet*, Université de Liège, Arlon Campus Environnement,](https://reader034.vdocuments.site/reader034/viewer/2022050413/5f89a5fb7e6dfa50393f8bca/html5/thumbnails/11.jpg)
MACSUR mid-term – Sassari – 1-4 April 2014
Results : modeling improvement Oensingen, blind run
Bias RMSE R²
NEE [gC m-2 day-1] -0.513 2.034 0.487
GPP [gC m-2 day-1] -0.307 2.392 0.737
RECO [gC m-2 day-1] -0.820 1.615 0.805
ET [mm day-1] -0.107 0.910 0.549
![Page 12: Bayesian inference of a dynamic vegetation model …...Bayesian inference of a dynamic vegetation model for grassland Julien Minet*, Université de Liège, Arlon Campus Environnement,](https://reader034.vdocuments.site/reader034/viewer/2022050413/5f89a5fb7e6dfa50393f8bca/html5/thumbnails/12.jpg)
MACSUR mid-term – Sassari – 1-4 April 2014
Results : modeling improvement Oensingen, after calibration (1000's of model outputs)
Blind runs [NEE or GPP, + montrer les coupes]
Calibrated runs (mode of the posterior distributions of parameters)
Bias RMSE R²
NEE [gC m-2 day-1] -0.013 1.991 0.477
GPP [gC m-2 day-1] 0.136 2.199 0.760
RECO [gC m-2 day-1] 0.123 1.260 0.808
ET [mm day-1] -0.020 0.675 0.751
[graphe avec bunch of data, NEE or GPP]
![Page 13: Bayesian inference of a dynamic vegetation model …...Bayesian inference of a dynamic vegetation model for grassland Julien Minet*, Université de Liège, Arlon Campus Environnement,](https://reader034.vdocuments.site/reader034/viewer/2022050413/5f89a5fb7e6dfa50393f8bca/html5/thumbnails/13.jpg)
MACSUR mid-term – Sassari – 1-4 April 2014
Conclusion
Bayesian sampling with DREAM_ZS :● Obtain a uncertainty assessment on model parameters● Obtain an interval on model output due to parameters
uncertainties● Assess model sensitivity to its parameters
Future work :● Interact with other modelers teams and intercompare...
![Page 14: Bayesian inference of a dynamic vegetation model …...Bayesian inference of a dynamic vegetation model for grassland Julien Minet*, Université de Liège, Arlon Campus Environnement,](https://reader034.vdocuments.site/reader034/viewer/2022050413/5f89a5fb7e6dfa50393f8bca/html5/thumbnails/14.jpg)
MACSUR mid-term – Sassari – 1-4 April 2014
Thanks for your attention
![Page 15: Bayesian inference of a dynamic vegetation model …...Bayesian inference of a dynamic vegetation model for grassland Julien Minet*, Université de Liège, Arlon Campus Environnement,](https://reader034.vdocuments.site/reader034/viewer/2022050413/5f89a5fb7e6dfa50393f8bca/html5/thumbnails/15.jpg)
MACSUR mid-term – Sassari – 1-4 April 2014
![Page 16: Bayesian inference of a dynamic vegetation model …...Bayesian inference of a dynamic vegetation model for grassland Julien Minet*, Université de Liège, Arlon Campus Environnement,](https://reader034.vdocuments.site/reader034/viewer/2022050413/5f89a5fb7e6dfa50393f8bca/html5/thumbnails/16.jpg)
MACSUR mid-term – Sassari – 1-4 April 2014
Results : modeling improvement Grillenburg, blind run:
Bias RMSE R²
NEE [gC m-2 day-1] -0.715 1.901 0.343
GPP [gC m-2 day-1] 1.242 2.556 0.664
RECO [gC m-2 day-1] 0.238 0.728 0.498
ET [mm day-1] 0.526 1.879 0.555
![Page 17: Bayesian inference of a dynamic vegetation model …...Bayesian inference of a dynamic vegetation model for grassland Julien Minet*, Université de Liège, Arlon Campus Environnement,](https://reader034.vdocuments.site/reader034/viewer/2022050413/5f89a5fb7e6dfa50393f8bca/html5/thumbnails/17.jpg)
MACSUR mid-term – Sassari – 1-4 April 2014
Results : modeling improvement Grillenburg, after calibration (1000's of modeled GPP):
Blind runs [NEE or GPP, + montrer les coupes]
Calibrated runs (mode of the posterior distributions of parameters)
Bias RMSE R²
NEE [gC m-2 day-1] -0.340 1.898 0.477
GPP [gC m-2 day-1] 0.465 1.991 0.741
RECO [gC m-2 day-1] 0.125 1.596 0.660
ET [mm day-1] 0.100 0.616 0.584
[graphe avec bunch of data, NEE or GPP]
![Page 18: Bayesian inference of a dynamic vegetation model …...Bayesian inference of a dynamic vegetation model for grassland Julien Minet*, Université de Liège, Arlon Campus Environnement,](https://reader034.vdocuments.site/reader034/viewer/2022050413/5f89a5fb7e6dfa50393f8bca/html5/thumbnails/18.jpg)
MACSUR mid-term – Sassari – 1-4 April 2014
Results : Error in measurements Homoscedastic :
Heteroscedastic :
![Page 19: Bayesian inference of a dynamic vegetation model …...Bayesian inference of a dynamic vegetation model for grassland Julien Minet*, Université de Liège, Arlon Campus Environnement,](https://reader034.vdocuments.site/reader034/viewer/2022050413/5f89a5fb7e6dfa50393f8bca/html5/thumbnails/19.jpg)
MACSUR mid-term – Sassari – 1-4 April 2014
The algorithm : DREAM_ZS ● Inverse problem:
● 12 parameters were sampled using 3 measurements variables: RECO, GPP, ET
● A multi-objective cost function (CF) was used : CF = f(RECO, URECO, GPP, UGPP, ET, UET)
● Uncertainties on measurement U were considered as homoscedastic or heteroscedastic (i.e., constant or variable):
→ HOMOSCEDASTIC:
→ HETEROSCEDASTIC:
Where U0 is a user-defined uncertainty for each variable X:
Optimal parameters=argmin (Observations−Modeled ( parameters))
U=U 0
U=U 0
2+
U 0
2 X∗X
X U0
RECO 1.5 gC m-2 day-1
GPP 3 gC m-2 day-1
ET 1 gC m-2 day-1