Evaluation of two European growth models for
Description: Evaluation of two European growth models for Douglas fir Journées CAQSIS, Montpellier, 5742016 1 Minna Pulkkinen, Holger Wernsdörfer François Ningre LERFoB, AgroParisTech, INRA, F-54000, Nancy, France Background Project Assessment of
Related Topics
Download Presentation
"Evaluation of two European growth models for" is the property of its rightful owner. Permission is granted to download and print the materials on this website for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Presentation Transcript
slide1. Evaluation of two European growth models for Douglas fir Journées CAQSIS, Montpellier, 5–7/4/2016 1 Minna Pulkkinen, Holger Wernsdörfer & François Ningre
LERFoB, AgroParisTech, INRA, F-54000, Nancy, France<br>
slide2. Background Project "Assessment of new forest production systems for Douglas fir – towards a simulation tool linking research, development and teaching"
General driving problematics: need to renew forest management
Increasing demand for wood as renewable material and source of energy
Requirement of sustainability, favourable carbon and nutrient balances, maintenance of biodiversity, ... 2<br>
slide3. Growth models as tools in development of new management scenarios
Key question: how well are models able to simulate new management scenarios?
Evaluation of model simulations against observed data: if a model is able to reproduce actualised contrasting management scenarios, it is likely to be able to simulate new scenarios as well
Evaluation data: repeatedly measured field experiments with (i) highly contrasting initial densities and (ii) varied timings and intensities of thinnings
Sensitivity analysis: Identifying relations and phases of development critical for model performance
model improvement 3<br>
slide4. Different types of growth models may have different ability to produce robust results outside their construction / parameterisation conditions
modelling paradigm: empirical, semi-functional, functional
growth unit and spatial explicitness
initial state and input data requirements
time step
incorporation of randomness in modelled processes
Aim of this study: evaluate two tree-based European growth models – fully empirical vs. semi-functional – against independent data in view of realistic simulation of new management scenarios for Douglas fir in France 4<br>
slide5. 5 Perin, Ligot et al. (2012)
For even-aged pure stands
of Douglas fir, Norway
spruce and two larch
species
Empirical, spatially
inexplicit
Parameterised with
extensive data sets
from S Belgium
Fundamental relationships:
(i) Hdom growth
(ii) tree circumference growth (as function of Hdom, stand density and tree size)
Mortality: maximum stand basal area w.r.t. mean tree circumference
Initial state at age t: plot size, stand age, tree diameters, site index parameter of Hdom vs. age curve (H50) (Perin 2013) Models: Gymnos<br>
slide6. 6 Models: SimCoP Ottorini (1991, 1995); based on work by Mitchell (1975)
For even-aged pure Douglas fir stands
Semi-functional, spatially explicit
Parameterised with very detailed tree analysis data from N-E France
Basic idea:
height growth (Hdom growth adjusted with vigour and foliar volume)
crown growth in 3D (restricted by neighbouring crowns)
“effective” foliar volume
stem volume growth
Mortality with (i) suppression by neighbouring trees and (ii) Reineke’s self-thinning rule
Initial state at age 0: tree coordinates, tree vigours, site index parameter of Hdomvs. age curve (H50)<br>
slide7. A subset of 62 plots from permanent field experiments in Baden-Württemberg, S-W Germany (FVA Freiburg)
Measurements at irregular time intervals
diameter on all trees at all observation time points
height on (a subset of) trees at some observation time points
cause and time of removal on all trees
exact information on mortality and thinnings
co-ordinates of trees not available in all plots 7 Evaluation data<br>
slide8. Plot selection criteria
pure one-storey Douglas fir stands
tree coordinates available
no interventions before first observation time point
Hdom known at least at three observation time points 8<br>
slide9. Initial state in Gymnos: tree diameters at first observation time point
measured values from data
social statuses of trees
drawn from theoretical distribution
Initial state in SimCoP: tree vigours (and coordinates) at age 0
estimated from data (h/Hdom) at first observation time point and adjusted for mean and variance of default theoretical distribution
social statuses of trees
drawn from theoretical distribution 9 Simulations: Initial state & parameterisation<br>
slide10. Site index parameter H50
estimated by fitting Hdom vs. age curve into measured data, conditional on default values of other parameters of curve
with SimCoP, fitting used also to select Hdom vs. age curve among three alternatives
Otherwise original parameterisation used without modifications 10<br>
slide11. Trees killed by natural hazards (wind, ice / snow, game animals, ...) included in thinnings
introducing additional, sometimes drastic thinnings in data
Timing of thinnings determined by age (cf. Hdom, RDI, ...)
easy comparison with observations
Intensity of thinnings determined by stocking density to be removed (cf. target density after thinning)
no artificial correction of possibly erroneous before-thinning mortality 11 Simulations: Thinnings<br>
slide12. Employing two thinning algorithms (Capsis)
List thinning
removing exactly same tree individuals as were removed in data
close imitation of observed thinnings, but difficult to use in further work (scenario simulations)
Size-class thinning
removing trees in diameter classes in same proportion as in data
selecting removed trees within class randomly
spatial control
fairly good imitation of observed thinnings, possible to use in further work 12<br>
slide13. Sources of randomness
Gymnos: generation of tree diameters at initial state
selection of trees in mortality process
SimCoP: generation of tree vigours at initial state
Size-class thinning algorithm: selection of trees within each class
With each simulation scheme (model + initial state generation + thinning algorithm) involving randomness 100 repeated simulations per plot 13 Simulations: Randomness<br>
slide14. 14<br>
slide15. Evaluation based on stand characteristics computed in a uniform manner from tree diameters and heights: stocking density, basal area, Hdom, Ddom
No extra variation between models due to different volume, biomass etc. equations
Output of each simulation scheme: 100 time series (repetitions) of stocking density, basal area, Hdom, Ddom in each of 62 plots
Need to summarise variation (i) over repetitions at each observation time point in each plot, (ii) over observation time points in each plot, and (iii) over plots 15 Evaluation criteria<br>
slide16. Average performance of each simulation scheme in each plot
compare mean simulated time series to observed time series
summarise resulting time series of relative errors into plot-wise diagnostics
(a) mean of relative errors over observation time points average systematic error over observation period ("relative bias")
(b) mean of absolute values of relative errors
aggregated measure of performance at individual observation time points
(c) standard deviation of relative errors
variation in performance between individual observation time points ("relative precision") 16<br>
slide17. Variation between simulated time series of each simulation scheme in each plot
compare every simulated time series to observed time series
summarise resulting set of relative errors into plot-wise diagnostics
(a) mean of absolute values of relative errors over observation time points
(b) standard deviation of relative errors
With Gymnos, first observation time point (having perfect match!) not included in diagnostics computation 17<br>
slide18. 18 Results: Average model performance in plot over simulation period Both models had slight difficulty to reproduce observed mortality
( overestimation of stocking density) Both models predicted Ddom well Both models overestimated basal area<br>
slide19. 19 Hdom: Gymnos predicted very well, SimCoP underestimated on average in majority of plots Small mean of relative errors over time does not necessarily mean good performance at individual time points!<br>
slide20. 20 Results: Effect of initial density Initial density affected average model performance: stronger overestimation of basal area with low densities, underestimation with high densities, best performance with medium densities (mostly present in parameterisation?)<br>
slide21. 21 Results: Importance of early performance Performance of SimCoP iat early phase of simulation affected its average performance over whole simulation period (in basal area and Ddom) If SimCoP was doing well at 1st observation time
point, it was doing well over the whole observation
period<br>
slide22. Both models were able to predict relatively well – with their original parameterisation – stand growth and thinning reaction of independent contrasting data
similar performance achieved with very different parameterisation effort
Starting from age 0, SimCoP was sensitive to early phase performance
Both models can be used for developing new management scenarios 22 Concluding remarks<br>
slide23. Acknowledgements 23 Access to growth models and help with implementation:
T. Bronner & J.-M. Ottorini (SimCoP)
J. Perin & G. Ligot (Gymnos)
Evaluation data: A. Albrecht
Capsis tools: F. de Coligny
Funding: Chaire “Forêts pour demain”, DGER, Région Lorraine, AgroParisTech, INRA
The UMR 1092 LERFoB is supported by a grant overseen by the French National Research Agency (ANR) as part of the “Investissements d'Avenir” program (ANR-11-LABX-0002-01, Laboratory of Excellence ARBRE)<br>
slide24. 24 Results: Effect of initial density Ddom Stocking density<br>
LERFoB, AgroParisTech, INRA, F-54000, Nancy, France<br>
slide2. Background Project "Assessment of new forest production systems for Douglas fir – towards a simulation tool linking research, development and teaching"
General driving problematics: need to renew forest management
Increasing demand for wood as renewable material and source of energy
Requirement of sustainability, favourable carbon and nutrient balances, maintenance of biodiversity, ... 2<br>
slide3. Growth models as tools in development of new management scenarios
Key question: how well are models able to simulate new management scenarios?
Evaluation of model simulations against observed data: if a model is able to reproduce actualised contrasting management scenarios, it is likely to be able to simulate new scenarios as well
Evaluation data: repeatedly measured field experiments with (i) highly contrasting initial densities and (ii) varied timings and intensities of thinnings
Sensitivity analysis: Identifying relations and phases of development critical for model performance
model improvement 3<br>
slide4. Different types of growth models may have different ability to produce robust results outside their construction / parameterisation conditions
modelling paradigm: empirical, semi-functional, functional
growth unit and spatial explicitness
initial state and input data requirements
time step
incorporation of randomness in modelled processes
Aim of this study: evaluate two tree-based European growth models – fully empirical vs. semi-functional – against independent data in view of realistic simulation of new management scenarios for Douglas fir in France 4<br>
slide5. 5 Perin, Ligot et al. (2012)
For even-aged pure stands
of Douglas fir, Norway
spruce and two larch
species
Empirical, spatially
inexplicit
Parameterised with
extensive data sets
from S Belgium
Fundamental relationships:
(i) Hdom growth
(ii) tree circumference growth (as function of Hdom, stand density and tree size)
Mortality: maximum stand basal area w.r.t. mean tree circumference
Initial state at age t: plot size, stand age, tree diameters, site index parameter of Hdom vs. age curve (H50) (Perin 2013) Models: Gymnos<br>
slide6. 6 Models: SimCoP Ottorini (1991, 1995); based on work by Mitchell (1975)
For even-aged pure Douglas fir stands
Semi-functional, spatially explicit
Parameterised with very detailed tree analysis data from N-E France
Basic idea:
height growth (Hdom growth adjusted with vigour and foliar volume)
crown growth in 3D (restricted by neighbouring crowns)
“effective” foliar volume
stem volume growth
Mortality with (i) suppression by neighbouring trees and (ii) Reineke’s self-thinning rule
Initial state at age 0: tree coordinates, tree vigours, site index parameter of Hdomvs. age curve (H50)<br>
slide7. A subset of 62 plots from permanent field experiments in Baden-Württemberg, S-W Germany (FVA Freiburg)
Measurements at irregular time intervals
diameter on all trees at all observation time points
height on (a subset of) trees at some observation time points
cause and time of removal on all trees
exact information on mortality and thinnings
co-ordinates of trees not available in all plots 7 Evaluation data<br>
slide8. Plot selection criteria
pure one-storey Douglas fir stands
tree coordinates available
no interventions before first observation time point
Hdom known at least at three observation time points 8<br>
slide9. Initial state in Gymnos: tree diameters at first observation time point
measured values from data
social statuses of trees
drawn from theoretical distribution
Initial state in SimCoP: tree vigours (and coordinates) at age 0
estimated from data (h/Hdom) at first observation time point and adjusted for mean and variance of default theoretical distribution
social statuses of trees
drawn from theoretical distribution 9 Simulations: Initial state & parameterisation<br>
slide10. Site index parameter H50
estimated by fitting Hdom vs. age curve into measured data, conditional on default values of other parameters of curve
with SimCoP, fitting used also to select Hdom vs. age curve among three alternatives
Otherwise original parameterisation used without modifications 10<br>
slide11. Trees killed by natural hazards (wind, ice / snow, game animals, ...) included in thinnings
introducing additional, sometimes drastic thinnings in data
Timing of thinnings determined by age (cf. Hdom, RDI, ...)
easy comparison with observations
Intensity of thinnings determined by stocking density to be removed (cf. target density after thinning)
no artificial correction of possibly erroneous before-thinning mortality 11 Simulations: Thinnings<br>
slide12. Employing two thinning algorithms (Capsis)
List thinning
removing exactly same tree individuals as were removed in data
close imitation of observed thinnings, but difficult to use in further work (scenario simulations)
Size-class thinning
removing trees in diameter classes in same proportion as in data
selecting removed trees within class randomly
spatial control
fairly good imitation of observed thinnings, possible to use in further work 12<br>
slide13. Sources of randomness
Gymnos: generation of tree diameters at initial state
selection of trees in mortality process
SimCoP: generation of tree vigours at initial state
Size-class thinning algorithm: selection of trees within each class
With each simulation scheme (model + initial state generation + thinning algorithm) involving randomness 100 repeated simulations per plot 13 Simulations: Randomness<br>
slide14. 14<br>
slide15. Evaluation based on stand characteristics computed in a uniform manner from tree diameters and heights: stocking density, basal area, Hdom, Ddom
No extra variation between models due to different volume, biomass etc. equations
Output of each simulation scheme: 100 time series (repetitions) of stocking density, basal area, Hdom, Ddom in each of 62 plots
Need to summarise variation (i) over repetitions at each observation time point in each plot, (ii) over observation time points in each plot, and (iii) over plots 15 Evaluation criteria<br>
slide16. Average performance of each simulation scheme in each plot
compare mean simulated time series to observed time series
summarise resulting time series of relative errors into plot-wise diagnostics
(a) mean of relative errors over observation time points average systematic error over observation period ("relative bias")
(b) mean of absolute values of relative errors
aggregated measure of performance at individual observation time points
(c) standard deviation of relative errors
variation in performance between individual observation time points ("relative precision") 16<br>
slide17. Variation between simulated time series of each simulation scheme in each plot
compare every simulated time series to observed time series
summarise resulting set of relative errors into plot-wise diagnostics
(a) mean of absolute values of relative errors over observation time points
(b) standard deviation of relative errors
With Gymnos, first observation time point (having perfect match!) not included in diagnostics computation 17<br>
slide18. 18 Results: Average model performance in plot over simulation period Both models had slight difficulty to reproduce observed mortality
( overestimation of stocking density) Both models predicted Ddom well Both models overestimated basal area<br>
slide19. 19 Hdom: Gymnos predicted very well, SimCoP underestimated on average in majority of plots Small mean of relative errors over time does not necessarily mean good performance at individual time points!<br>
slide20. 20 Results: Effect of initial density Initial density affected average model performance: stronger overestimation of basal area with low densities, underestimation with high densities, best performance with medium densities (mostly present in parameterisation?)<br>
slide21. 21 Results: Importance of early performance Performance of SimCoP iat early phase of simulation affected its average performance over whole simulation period (in basal area and Ddom) If SimCoP was doing well at 1st observation time
point, it was doing well over the whole observation
period<br>
slide22. Both models were able to predict relatively well – with their original parameterisation – stand growth and thinning reaction of independent contrasting data
similar performance achieved with very different parameterisation effort
Starting from age 0, SimCoP was sensitive to early phase performance
Both models can be used for developing new management scenarios 22 Concluding remarks<br>
slide23. Acknowledgements 23 Access to growth models and help with implementation:
T. Bronner & J.-M. Ottorini (SimCoP)
J. Perin & G. Ligot (Gymnos)
Evaluation data: A. Albrecht
Capsis tools: F. de Coligny
Funding: Chaire “Forêts pour demain”, DGER, Région Lorraine, AgroParisTech, INRA
The UMR 1092 LERFoB is supported by a grant overseen by the French National Research Agency (ANR) as part of the “Investissements d'Avenir” program (ANR-11-LABX-0002-01, Laboratory of Excellence ARBRE)<br>
slide24. 24 Results: Effect of initial density Ddom Stocking density<br>