Sunteți pe pagina 1din 108

The Baltic Sea cod collapse:

a social-ecological perspective
Steven J. Lade, K. Orach, Maja Schlter and lots more!
Natural Resource Modelling conference
Vilnius, July 2014

Baltic Sea
social system
(fisherman
behaviour,
markets,
regulation, )

Overfishing
Abiotic drivers

Baltic Sea
Ecosystem

Cod collapse

(salinty, temperature,
nutrients, oxygen)

What role might social processes have played in the collapse of cod?
Might changes in human behaviour have contributed to, or alternatively held
back, the collapse?
What lessons can we learn for future governance of the Baltic Sea? How can
we avoid similar regime shifts in the future?

Outline
Social-ecological conceptual framework
Modelling approach
1. Conceptual: Our particular S-E model
2. Technical: Generalized modeling
Preliminary results

Social-ecological modelling approach


Ecological Modelling

Schlter et al., Natural Resource Modeling, 2012

Bioeconomic Modelling

Social-ecological Modelling

Kirill Orach

Maja Schlter

Steven Lade

Stockholm Resilience Centre Baltic Sea experts

Thorsten
Blenckner

Martin Quaas, U Kiel

Susa
Niiranen

Jonas
HentatiSundberg

Wijnand
Boonstra

Henrik
sterblom

Causal loop diagram of pre-collapse SES


HerringFishing
-

immigration

AccessRegulations

+
External
Fleet

OpportunityCost

Local Fleet Size

+
+

TotalFleetSize
+
FishingEffort
+

LocalFleet

Sunk
Cost
+
+
Effects

+
+

TimeSpent
Fishing

Social
Identity

Technological
Development
+
+
CodCatch +

Oxygen

+
CodBiomass ++
(adult) +

+
Subsidies

Subsidies

EcologyCodLarvae +
Biomass
(cod, juvenile cod,
+ sprat,
Salinity
+ CodEggs +
herring, Pseudocalanus)

Benthos

+
+ CodJuvenile
+
Biomass
+

ExpectedMarginalCosts

MarginalProfit
Perception

Herring
Biomass

+
Delayed
CPUE
perception
Perception
+
+
of cod
stock
MarginalReturn
+
Perception

Sprat
Biomass

Price +

ExternalCodSupply

Cod market

ConsumerPreferences
ForCod

How to make a dynamical systems model??

Temperature
SpratFishing

Pseudocalanus
Biomass
+

Generalised modelling
Analyse dynamical system for stability and bifurcations without specifying
complete functional forms
Rapidly explore an entire class of models
Successful applied in ecology, molecular biology, bone physiology, socialecological systems

Generalized modeling: example


Single population X with reproduction G(X) and mortality H(X)

1. Write down generalised model:

= ()

2. Assume there exists a fixed point: X = X*


3. Symbolically calculate Jacobian: =

4. Re-write in terms of generalized parameters: = ( )


=

characteristic time scale of state variable

Point elasticities: =

nonlinearity of link

parameters: ratio of link strengths


5. Calculate eigenvalues: = ( )
6. Parameter variation: If linear mortality = 1 and
0 < < 1 (e.g. Holling type II) always stable
0 < < 2 (e.g. Holling type III) may become unstable

Causal loop diagram of pre-collapse SES


HerringFishing
+
+

TotalFleetSize
+

AccessRegulations

FishingEffort
LocalFleet

External
Immigration
Fleet
OpportunityCost

Sunk Cost
+ TimeSpent
Fishing
Effects
+

Technological
Development
+
+
CodCatch +

Subsidies

Delayed
+
perception
CPUE
of +codPerception
stock

Oxygen

+
CodBiomass ++
(adult) +

EcologyCodLarvae +
Biomass
(cod, juvenile cod,
sprat,
+
Salinity
+ CodEggs +
herring, Pseudocalanus)

Benthos

+
CodJuvenile
+
+
Biomass
+

ExpectedMarginalCosts

Subsidies

MarginalProfit
Perception

MarginalReturn +
Perception

Sprat
Biomass

Price +

ExternalCodSupply

Economic
ConsumerPreferences
feedbacks
ForCod
+

How to make a dynamical systems model??

Herring
Biomass

Focus on a single event: cod collapse = Decreased stability


Modelling objective: Calculate stability of the dynamical system
Use generalised modelling

Temperature
SpratFishing

Pseudocalanus
Biomass
+

Data for generalised parameters


Ecological
Biomass flows from ICES data
Predation elasticities from existing models
Social
Fleet composition
Subsidies: fleet purchase and income support
Economic: Consumption, standard models
Immigration: change in regulations
But no data on fisherman psychology

13

Preliminary Results
Stability

Direction of collapse

(dominant eigenvalue)

(eigenvectors)

Unstable

Stable

Model validation: increasing instability


towards start of collapse
Model analysis: roles of subcomponents

Model validation:
Cod, JuvenileCod, Pseudocalanus decrease
Sprat, TimeSpentFishing increase

Correlated with instability

Sensitivity of stability to individual parameters

Fast: Over 80 parameters but only need eigenvalues


Model validation: Sensitivity to assumptions
Model analysis: Leverage points

Anticorrelated with instability


(correlated with stability)

Loop dominance
What is the relative contribution of each feedback loop to overall stability?
In systems dynamics, cumbersome procedure
Directly from Jacobian:
use Leibniz formula
1 11

det =
1, ,

where 1 is the Levi-Civita symbol,


along with det = 0 at fold bifurcation

Which feedback loops were dominant?


(contribution of terms to the determinant)

Model experiments
Sensitivity to assumptions about fisherman behaviour
Profit-maximising vs satisficing
Hypotheticals: What if
Subsidies were not present?
There had been effective monitoring and regulation? Scientific advice
had been heeded?
Fisherman could have easily switched to other fish?
Ecology had been different?
West-coast fishermen had not come in to the Baltic?
Question: Other processes it might be worthwhile exploring?

Summary
HerringFishing

AccessRegulations
+

+TotalFleetSize

External
Fleet

+
+

FishingEffort

Technological
Development

LocalFleet
+

Herring
Biomass

+
+
CodCatch +

TimeSpent
Fishing
+

Oxygen

+
CodBiomass ++
(adult) +

Benthos

+
+ CodJuvenile
+
Biomass
+

CodLarvae +
Biomass
+
+ CodEggs +

Salinity

Pseudocalanus
Biomass
-

Subsidies

+
ExpectedCosts

+
PerceivedProfitability
+

+
ExpectedIncome +

+
Perceived
CodStock

Sprat
Biomass

Price +

ExternalCodSupply

Temperature
SpratFishing

ConsumerPreferences
ForCod

Social-ecological Modelling

&

Exploring the roles of linkages and feedbacks in


social-ecological systems using theoretical models
and grounded case studies

Statistical concepts and methods in


forest fire history studies
William J. Reed
Department of Mathematics and Statistics
University of Victoria
Victoria B.C.
Canada V8W 3P4.
www.math.uvic.ca/faculty/reed/index.html

Outline
Existing fire history concepts and definitions ambiguities and contradictions. Revised definitions.
Estimating historical fire interval from scar data (surface fires); and
Estimating historical fire interval from time-sincefire-map data (stand-replacing fires).
Also testing for change points for fire frequency; and
identifying change points.

Definitions currently in use (Johnson & Van Wagner, 1985; Johnson & Gutsell, 1994):
hazard of burning - statisticians hazard rate at any
location.
fire interval - exepected time between fires at a point
(reciprocal of hazard of burning if assumed constant
over time).
annual percent burn
fire cycle - time to burn an area equal to that of
the study area (reciprocal of annual percent burn if
constant from year to year).

It is common practice to use the notions of fire cycle


and fire interval interchangeably.
This arises from a deterministic way of thinking (so that
annual percent burn is thought of as equal to probability
of a fire at any location).
Statisticians will recognize that the annual percent burn
is a random variable; in consequence the time to burn
an area equal to that of the study area, is also a random
variable. So fire cycle not well defined!
Identifying the fire interval with the time to burn an area
equal to that of the study area, has caused confusion
especially in simulation studies.

Revised definitions.
Assume temporal homogeneity for the fire process over
a given area (size A) and define the area-wide hazard
of burning as
"

P (fire somewhere in study area in (t, t + )


= lim
.
0

(independent of t because of assume temporal homogeneity). Now define the local hazard of burning
"

P (fire at location x in(t, t + )


.
0

(x) = lim

Let h(x, y) be the conditional probability that a fire ignited at location x spreads to location y. If the location
y of an ignition has a probability density f (y) over the
area A, it follows that
(x) =

Z
A

h(x, y)f (y)dy = p(x),

say.

In the absence of detailed spatial information, assume


spatial homogeneity, so that
p(x) p

for all locations x

where p denotes the conditional probability of a fire at


any given location given a fire somewhere in the study
area. Thus
(x) = p

for all locations x


6

The fire interval can be defined as the expected time


between fires at any location. So

FI =

1
1
=

(One could also define an area-wide fire interval expected time between fires anywhere in study area 1 = pF I.)
equal to

The notion of fire cycle is not well-defined and should


be abandoned.
Instead use notions of local hazard of burning and its
reciprocal fire interval to characterize fire frequency.

14

ESTIMATION OF HISTORICAL FIRE


FREQUENCY FROM SCAR DATA.

Assume a simple random sample of trees has been examined for scars.

Does this ever happen in practice?

15

Example of simplified scar data (fake).

There are 5 sample trees; 7 distinct fires; and 14 scars.


The period 1890-2000 can be divided into 6 epochs (with constant
numbers of trees).

Preliminaries and notation.


First divide past time into distinct epochs during
each of which a constant number of sample trees are
vulnerable to fire.
(The break points between epochs are when sample
objects arrive or leave).

17

In epoch j (= 1, 2, . . . , J) let

Nj denote number of trees in sample vulnerable to


scarring;
nj denote number of distinct dates at which fires were
recorded;
tj,r (r = 1, . . . , nj ) denote times (since start of epoch)
at which fires were recorded;
xj,r (r = 1, . . . , nj ) denote number of scars registered
for fire i and
n

j
xj. = Pr=1
xj,r denote total number of scars registered in epoch.
18

Also let

mj denote the number of dates when fires which registered no scars on the sample trees were ignited in
epoch j.
N.B the mj are unobserved variables.
The total number of fires ignited in the area over the
total time T (since start of first epoch) is
J
X

nj + mj = n + m

j=1

19

We wish to estimate model parameters (area-wide


hazard of burning) and p (conditional probability of a
fire at any location given a fire somewhere.)
The full-data likelihood is the probability of such observations (including missing values mj , j = 1, . . . , J
- the numbers of fires leaving no scars).

20

Now if scars were registered independently of one another on sample objects, given that a fire has ignited,
the distribution of the number of scars registered (Xj,r )
for a given fire (r) in epoch j would be
Bin(Nj , p)
But clearly assumption of independence not valid (fires
spread spatially!) as following results of a binomial dispersion test (for Dugout region of Blue Mountains, E.
Oregon) indicate:

21

To allow for contagion in spread of fires consider instead


an overdispersed binomial model with pmf f (x; N, p, )
where is overdispersion parameter. So that (with
q = 1 p)
x log(p/q) + log q N
lnf (x; N, p, ) =
+ c(, x)

and use quasi-likelihood for estimation.

22

Under these assumptions the (full data) quasi-likelihood


is of the form
L(, p; ) = km.+n. eT

J
Y

f (0; Nj , p, )mj

j=1

NB. this includes missing data (mj ,

nj
Y

f (xj,r ; Nj , p, )

r=1

j = 1, . . . J).

The log quasi-likelihood is thus

Q(, p; ) = c + (n. + m.) log T


i
X
1h X
Nj nj x..) log(1 p) + x.. log p
+ ( Nj mj +

23

MLEs of and p can be found by maximizing quasilikelihood over and p using EM algorithm;
to estimate overdispersion parameter use a moment
estimator.
It turns out that EM can be done analytically!

24

The M-step and E-step together yield the following


(coupled) recuurence equations relating the estimates
at i + 1 step, with those at i step:

(i+1) =

p(i+1) =

(i)
n. +

(1 p(i))Nj j
T

x..
(i)

P
N
(i)
j
Nj j (1 p ) + Nj nj

where j is length of epoch j.

25

The coupled system converges to an equilibrium which


and q = 1 p
yields the MLEs
=

1 q =

n.
T

P N
q j j

x..(T
n.

Nj j qNj +

P N
q j j )

Nj nj (T

j qNj )

N.B. Second equation has to be solved numerically.


The MLE of the fire interval is
FI =

1
(1 q)

26

and q can be found from


Asymptotic variances of
information matrix (and moment estimate of overdispersion parameter, ) , and variance of FI estimated
as
)
var(
q)
var(

2

+
var(
F I) = F I
2

q2
"

27

Fake data.

MLEs:

= 0.0619;

q = 0.4006;

FI = 24.49 years (s.e. = 10.66)

95% CI for FI: 3.6 45.4 years. Estimate of overdispersion parameter = 1.971
NB. Mean (and standard deviation) of 9 inter-fire intervals is 25.22 (and 20.74) years.

Example 2. Arroyo de las Flores, Mexico (E. Heyerdahl


& Alvarado,E.)
Data has 22 sample trees (from 1779 to 1995); 26
epochs; 43 fires and 236 scars.
The MLE of fire interval is FI = 14.18 years with a
standard error of 2.63 years. (Estimate of OD parameter = 4.81).

29

ESTIMATION OF HISTORICAL FIRE


FREQUENCY FROM TIME-SINCE-FIRE DATA.
Used for montane and boreal areas with stand-replacing
fires.
For time-since-fire map data the area is partitioned
into non-overlapping regions according to the time since
last fire (essentially stand age).
Thus one has a complete record of the extent of the
most recent fire but for earlier fires only an incomplete
record (since the area of an earlier fire may have been
burned over, in part, by more recent fires).
For time-since-fire sample data the time since last
fire at a sample of points (chosen according to welldefined sampling plan) is observed .
N.B. Usually data are binned into 10 or 20 year age
classes,
30

Example of time-since-fire map data


Boundary Waters, MN Heinselman,1973.

Example of time-since-fire sample data


Wood Buffalo N.P, AltaNWT (Larsen, 1996)

Divide total time (age of oldest stands) into distinct


epochs, separated by change points.
In each epoch the hazard of burning is assumed constant.
How to do this?
Informal method: see where semi-log cumulative plot
seems to break into line segments + expert knowledge.
More formal method: use model identification methods (e.g. Bayes Information Criterion)

33

Theory:
again assume homogeneity over study area.
at any given location the probability j that it belongs to time-since-fire class j (i.e. originated between
(j 1)T and jT years ago) can be calculated as the
probability of a fire at the location between (j 1)T
and jT years ago; and no fire there subsequently. Thus

j = exp

j1
X

h
i
(i)T
(i)
(1 exp T ) )

i=1

where (i) is the (local) hazard of burning which prevailed between (i 1)T and iT years ago.
34

e.g. if there are no change points (one epoch) with


constant hazard ( )
j = e(j1)T (1 eT )
and if there are two epochs with one change point
(pT years ago say) with hazards 1 and 2 prevailing
after and before, then
j = e(j1)1T (1 e1T ),

for j = 1, 2, . . . , p

and
j = ep1T e(jp1)2T (1e2T ), for j = p + 1, p + 2, . . .
and so on for two or more change points ....

If fires didnt spread spatially one could assume independence of times since fire at distinct locations, and
thus a multinomial distribution for areas in age classes;
but because of contagious effects of fire independence is again unrealistic; instead assume an overdispersed multinomial distribution and corresponding
quasi-likelihood:
1X
Q=
yj log j
j
where is overdispersion parameter; and yj is either the
number of sample points, or the area in time-since-fire
class j
Can maximize over 1 (for single epoch model); or over
1 and 2 (for two epoch model); etc........
35

Standard errors can be computed from the inverse of


information matrix.
Confidence intervals by Wald or likelihood ratio methods.
Significance of change point can be tested by LR
methods.

36

Two epoch model fitted to Wood Buffalo data


Change point at 1860
MLEs of FI are 70 yrs (CI: 58-90) for 1861-1989; and 34 yrs (CI: 23-69) 1860 and earlier.

Do we need more change points?


If so, how do we determine them?

Determining number and location of change points is


a problem in model identification.
One possibility is to use Bayes Information Criterion
BIC.
Here applied to two data sets:
Kananaskis Watershed (Johnson & Larsen, 1991) and;
Glacier National Park (Johnson, Fryer & Heathcott,
1990).

38

Example of time-since-fire map data


Kananaskis watershed Johnson & Larsen.
Age intervals of width 10 years.

Model
H0
H1
H2
H3
H4
H5
H6

MLEs of change points


4
4, 6
4, 6, 24
4, 6, 13, 23
4, 6, 7, 13, 23
4, 6, 7, 13, 19, 27

BIC
-55.52
-149.06
-156.35
-159.39
-153.35
-149.42
-136.71

Posterior probability
0.000
0.004
0.171
0.780
0.038
0.005
0.000

Maximum likelihood estimates of change points in various models,


along with associated BICs and posterior probabilities of the various models, assuming a priori that all seven models are equally
probable (for Kananaskis Watershed time-since-fire data).

The three change point (4 epoch) model seems most


plausible
40

Glacier National Park Johnson, Fryer & Heathcott.


Age intervals of width 20 years.

Model
H0
H1
H2
H3
H4
H5
H6

MLEs of change points


16
12, 16
5, 10, 16
2, 5, 10, 16
2, 5, 10, 12, 16
2, 5, 10, 12, 16,18

BIC
201.63
74.38
42.79
28.08
-9.27
-6.98
1.76

Posterior probability
0.000
0.000
0.000
0.000
0.797
0.202
0.002

Maximum likelihood estimates of change points in various models,


along with associated BICs and posterior probabilities of the various models, assuming a priori that all seven models are equally
probable (for Glacier National Park time-since-fire data).

The four change point (5 epoch) model seems most


plausible.

41

Kananaskis
Watershed

Glacier
National Park

Epoch
i
1
2
3
4
Epoch
i
1
2
3
4
5

Date
1940 - 1980
1920 - 1940
1740 - 1920
pre 1740
Date
1940
1880
1780
1660
pre

- 1980
- 1940
- 1880
- 1780
1660

Fire Cycle (years)


MLE
95% Con. Int.
6409
969 715, 000
49
34 73
136
101 189
48
30 85

Fire Cycle (years)


MLE
95% Con. Int.
1980
565 16700
156
40 181
1827
673 8102
151
106 224
25
17 42

Maximum likelihood estimates and 95% LR confidence intervals


for the fire cycles between the estimated change points in best
models.
42

Epochs and corresponding hazards


as determined from BIC

References.
Concepts: Reed, W. J. A note of fire history concepts and definitions. To appear Can. J. For. Res.
Fire scars: Reed, W.J. and Johnson, E. A. 2004.
Statistical methods for estimating historical fire frequency from multiple fire-scar data. Can. J. For. Res.
34, 2306-2313.
Time-since-fire data: Reed, W.J., Larsen, C.P.S.,
Johnson E.A. and MacDonald, G.M. 1998. Estimation
of temporal variations in historical fire frequency from
time-since fire map data. For. Sci. 44: 465-475.
Model selection: Reed, W.J. 1998. Determining
changes in historical forest fire frequency from a timesince-fire map.J.Agri Biol. Environ Stat 3: 430-450.
Reed, W.J. 2000. Reconstructing forest fire history
identifying hazard rate change points using the Bayes
Information Criterion. Can. J. Stat. 28: 352-365.
44

Using the viability theory for evaluating management policy of


forest systems under environmental intensification
Jean-Denis MATHIAS1, Bruno BONT2, Thomas CORDONNIER3, Francis DE MOROGUES4

1IRSTEA

Clermont-Ferrand, Laboratoire d'Ingnierie pour les Systmes


Complexes (LISC),
2IRSTEA

Montpellier, Gestion de l'Eau, Acteurs, Usages (GEAU),

3IRSTEA

Grenoble, cosystmes Montagnards (EM),

4Fort

Cellulose Bois-construction Ameublement (FCBA)

1
MATHIAS J-D. et al

WCNRM'14, Vilnius

Context
Multifunctionnality of forest

http://www.arange-project.eu/

Focus on two main services in the case of uneven-aged forest


Timber production

Biodiversity

http://microcomputer/photos/undpeuropeandcis/8099867489/

http://actionbiodiversity.org/2010/07/human-benefits-of-biodiversity/biodiversity-forestlarge/

Socio-economic issues

Societal and environmental issues

Trade-off management
2

MATHIAS J-D. et al

WCNRM'14, Vilnius

Context
Forest multifunctionality management
The forest is a (partially) controled system!
Environmental
recommendation
Preservation function

Timber demand
Production function

Need for mathematical tools for managing forest multifunctionality


3
MATHIAS J-D. et al

WCNRM'14, Vilnius

Context
Applying viability theory
Identify and qualify the set of viable states
X2

X1

4
MATHIAS J-D. et al

WCNRM'14, Vilnius

Context
Applying viability theory
Identify and qualify the set of viable states
X2

X1

5
MATHIAS J-D. et al

WCNRM'14, Vilnius

Context
Applying viability theory
Identify and qualify the set of viable states
X2

X1

6
MATHIAS J-D. et al

WCNRM'14, Vilnius

Context
Applying viability theory
Identify and qualify the set of viable states
X2

X1

7
MATHIAS J-D. et al

WCNRM'14, Vilnius

Context
Applying viability theory
Identify and qualify the set of viable states
X2

X1

8
MATHIAS J-D. et al

WCNRM'14, Vilnius

Context
Applying viability theory
Identify and qualify the set of viable states
X2

X1

9
MATHIAS J-D. et al

WCNRM'14, Vilnius

Context
Applying viability theory
Identify and qualify the set of viable states
X2

K
Viability
kernel
ker(K)

Flexibility

Local

X1

Global

f x =

F (,K )=

uv x
U

f (k )dk

dk

10
MATHIAS J-D. et al

WCNRM'14, Vilnius

Context
Applying viability theory
for studying forest practices under environmental intensification
Environmental
preservation
Increase of the volume of the
deadwood in the stand

Timber production
Forest dynamics + Viability Problem
Model

Increase of the
timber demand

Which impact has the intensification of


both functions on the flexibility?

FLEXIBILITY CHANGE

11
MATHIAS J-D. et al

WCNRM'14, Vilnius

Outline

I. Model and viability problem


II. Scenarios and results
III. Conclusions

12
MATHIAS J-D. et al

WCNRM'14, Vilnius

I Model and viability problem


Two strata forest dynamics model
Inspired from : T.S. Kohyama and T. Takada., 2012
Recruitment

Competition

Growth
Mortality
Decomposition
13
MATHIAS J-D. et al

WCNRM'14, Vilnius

I. Model and viability problem


Two strata forest dynamics model
+ Cuts and Timber stock management
Recruitment

Competition

Growth

Removal
Wood processing
industry

Mortality
Decomposition
14
MATHIAS J-D. et al

WCNRM'14, Vilnius

I. Model and viability problem


Two strata forest dynamics model
+ Cuts and Timber stock management
r

Recruitment

Competition

Growth

Removal
Wood processing
industry

Mortality
MATHIAS J-D. et al

Decomposition

15
WCNRM'14, Vilnius

I. Model and viability problem


Two strata forest dynamics model
+ Cuts and Timber stock management
r

16
MATHIAS J-D. et al

WCNRM'14, Vilnius

I. Model and viability problem


Two strata forest dynamics model
+ Cuts and Timber stock management
r

17
MATHIAS J-D. et al

WCNRM'14, Vilnius

I. Model and viability problem


Two strata forest dynamics model
+ Cuts and Timber stock management
r

18
MATHIAS J-D. et al

WCNRM'14, Vilnius

I. Model and viability problem

The constraint set K


Minimum dead wood volume
(Environmental recommendation)

Vdmin < Vd < Vdmax

Positive timber wood stock

0<W

Positive number of mature trees

0 < X1

Positive number of Juveniles

0 < X2

19
MATHIAS J-D. et al

WCNRM'14, Vilnius

I. Model and viability problem

The constraint set K


Minimum dead wood volume
(Environmental recommendation)

Vdmin < Vd < Vdmax

Positive timber wood stock

0<W

Positive number of mature trees

0 < X1 < X1max

Positive number of Juveniles

0 < X2< X2max

20
MATHIAS J-D. et al

WCNRM'14, Vilnius

I. Model and viability problem

Intensification of the production


function:

parameter

Intensification of the production


function:

Forest dynamics + Viability Problem


Model

Constraint Vdmin

Which impact has the intensification of


both functions on the flexibility?

FLEXIBILITY CHANGE

21
MATHIAS J-D. et al

WCNRM'14, Vilnius

Outline

I. model and viability problem


II. Scenarios and results
III. Conclusions

22
MATHIAS J-D. et al

WCNRM'14, Vilnius

II Scenarios and results


Tested scenarios

Scenarios
Constraint Vdmin
Parameter

from

15

to

35 m3/ha

from

to

14 m3/y/ha

Calibrated from a real forest, located near Grenoble (France)

23
MATHIAS J-D. et al

WCNRM'14, Vilnius

II Scenarios and results


!

4-D kernel!
W=45m3

Local
flexibility

f x =
Vd

uv x
U

Tested scenario:

8 m3/y/ha
Vdmin
20 m3/ha

Local flexibility
-

<

24
MATHIAS J-D. et al

WCNRM'14, Vilnius

Global flexibility

II Scenarios and results

Local
flexibility

f x =
Wood stock (m3)

uv x
U

8 m3/y/ha
Vdmin
20 m3/ha

25
MATHIAS J-D. et al

WCNRM'14, Vilnius

II Scenarios and results

Global
flexibility
Vdmin

F (,K )=

f (k )dk

dk

Area with high flexibility


26
MATHIAS J-D. et al

WCNRM'14, Vilnius

Outline

I. Model and viability problem


II. Scenarios and results
III. Conclusions

27
MATHIAS J-D. et al

WCNRM'14, Vilnius

III Conclusions
Take home messages
- We consider a simple model of forest management
- A flexibility indicator has been developed using the
viability theory for helping managers in their decision
making

- Flexibility strongly decreases from 30 m3/ha of


deadwood and a timber demand equal to 10 m3/y/ha

28
MATHIAS J-D. et al

WCNRM'14, Vilnius

III Conclusions

Thank you for your attention!

More details

https://forgeco.cemagref.fr
29
MATHIAS J-D. et al

WCNRM'14, Vilnius

Generalized HeightDiameter Models with


Mixed-Effects Parameters
Diffusion Processes
Petras Rupys

Aleksandras Stulginskis University, Lithuania

Need Height Diameter,


Crown width Diameter,
Stem Profile Models !

Current Models

Linear and non-linear regression models:

Generalized models
Random effects parameters models

Currently used regression methods are not


defendable statistically, because
both
assumptions of non-normality and nonhomogeneity of the variance of errors
distribution are more often than not rejected

Deterministic and Stochastic!


The deterministic (Mitcherlich)
dy x
y (x)
dx
y ( x0 ) y0

The stochastic (Vasicek)

dY x Y ( x) dx dW x

Solution - non-sigmoidal curve


y (x)

Solution - density function


for each x
p ( y, x)

P(Y ( x0 ) y0 ) 1

Probability density function of


the total tree height

New Stochastic Differential


Equation Model

The stochastic differential equation of the height dynamics, Y,


versus diameter, x, is defined in the following Vasicek form

dY x Y ( x) dx dW x
where: x 0; X 0 , PY 0 1.3 1 , , ,
parameters - to be estimated from data.

- fixed effects

Generalizations of Stochastic
Vasicek Model

The parameter has high variation between stands. So:


random effects are introduced
0 i

where i~ N 0; (i=1, 2, ..., M) are stand-specific


random effects, M is the number of measured stands
exogenous stand variables (k=1,2,,q) are introduced by
2

g ( x) 0 k g k ( x)
k 1

where g k (x) are polygonal functions defined by


i
i
y

y
g k ( x) y0i k jk i 0 k x; 0 x x ijk , j 1,2,..., ni
x jk

Generalized Mixed Effects


Stochastic Differential Equation
Model
q

dY x 0 i k g k ( x) Y i ( x) dx dW i x
k 1

PY i 0 1.3 1

x 0; X 0

The process Yi(x) corresponds to a normal distribution

N i ( x); 2 ( x)
where

( x) 1.3e
i

d
q
0 i
x
1 e k g k (u)e xu d x

k 1
0

2 x
1

e
2 ( x)
2
2

Parameters Estimation of
Vasicek type SDE

The log-likelihood function :


for generalized fixed effects parameters model is defined by
LL1 ( 0 ,1 ,..., q , , ) ln f y ij , x ij
M

ni

i 1 j 1

where
2
x
q

x
x
x u
0
y 1.3e 1 e k g k (u )e
dx

k 1

1
0

f y, x
exp

22 x
22 x

for generalized mixed-effects parameters


model

LL2 ( 0 , 1 ,..., q , , , ) ln f ( y ij , x ij ) ln p(i ) di


M

ni

i 1 R j 1

where

2
x
q

x
x
x u
0
i
y 1.3e
1 e k g k (u )e d x

k 1

1
0

f y, x
exp
2

2 x
22 x

i 2
p(i )
exp 2
2
2
2
1

Maximum Log-likelihood of
Generalized Mixed Effects Model
The Laplace method is used, let

g i 0 ,1 ,..., q , , , ln f ( y ij , x ij ) ln p(i )
ni

j 1

i arg max g i 0 ,1 ,..., q , , ,

2 g i 0 ,1 ,..., q , , ,
i
i
H ( i 0 ,1 ,..., q , , , )
2i

LL2 ( 0 ,1 ,..., q , , , )

1
1

g i 0 ,1 ,..., q , , , ln 2 ln H ( i 0 ,1 ,..., q , , , )

2
i 1
M

Parameter Calibration

If a sub-sample of m trees with height yi and diameter xi, i=1, 2,


..., m, is taken from a new stand, the random-effects parameter
for the new stand can be calibrated in the following form

k z0 k 1 exp( x j )

1 q

y j 1.3 exp( x j )

z
z

jk
0k

k 1 z z

1 exp( x j )
jk
0k


xj
1 m

m j 1

1 exp( x j )

Stochastic Predictions

ystoch

( x); 2 ( x)

1
U

U is a uniform random variable in the interval


lowermost 5%
uppermost 5%

0.05; 0.95

Data

The number of pine trees of the estimation dataset 1627 (12 stands)
The number of pine trees of the validation dataset 699 (5 stands)

Generalized fixed effects (left) and mixed-effects (right) parameters height models for all
validation datasets using exogenous stand variables: without (solid - black), crown
height (dot - blue), mean height (dash - green), mean diameter (dot dash - gold)

Variables Selection and


Model's Comparison
oAkaikes Information Criteria

oByasian Information Criteria

AIC 2 LL ( ) 2 p

BIC 2 LL ( ) ln( n) p

oTest - the ratio of the maximum log likelihoods of the models

LL ( 2 )
T 2 ln 2

LL

(
)
1
1

where statistic T is asymptotically distributed as a chi-squared


random variable with degrees freedom p2-p1, where p1 and p2
are the number of free parameters of models 1 and 2

P-values of the ratio test T

stocking level, S

soil fertility index, BAR


crown width, cW
age, A
mean height, h0
mean diameter, d0
crown height, cH

Conclusions
1.

2.

3.

The Vasicek, Gompertz and von Bertalanffy


shape stochastic processes were utilized to
model the total tree height, crown width and
stem taper.
The maximum likelihood estimators for fixed
effect and mixed effects parameters models
were derived.
The developed variance functions can be
applied for generating weights in linear and
nonlinear weighted least squares regression
models.

References
1. Rupys, P., Petrauskas, E. 2010. The Bivariate Gompertz Diffusion

Model for Tree Diameter and Height Distribution. Forest Science,


56: 271-280.
2. Rupys, P., Petrauskas, E. 2010. Quantyfying Tree Diameter
Distributions with One-dimensional Diffusion Processes. Journal of
Biological Systems, 18: 205-221.
3. Rupys, P., Petrauskas, E. 2012. Analysis of Height Curves by
Stochastic Differential Equations. International Journal of
Biomathematics, 5(5), DOI No: 10.1142/S1793524511001878.
4. Rupys, P. 2013. The Further Development of Stem Taper and
Volume Models Defined by Stochastic Differential Equations.
Lecture Notes in Electrical Engineering, 229: 121-134, Springer.
5. Rupys, P. 2014. Height-Diameter Models with Stochastic Differential
Equations and Mixed Effects Parameters. Journal of Forest
Research, 19: DOI 10.1007/s10310-014-0454-1.

Thank for Attention!

Questions and Comments

S-ar putea să vă placă și