improving the accuracy of predicted diameter and height distributions jouni siipilehto finnish...

22
Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: [email protected]

Upload: may-richards

Post on 04-Jan-2016

215 views

Category:

Documents


2 download

TRANSCRIPT

Page 1: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

Improving the accuracy of predicted diameter and

height distributions

Jouni SiipilehtoFinnish Forest Research Institute, Vantaa

E-mail: [email protected]

Page 2: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

Introduction

• Diameter distributions are needed in Finnish forest management planning (FMP)– individual tree growth models

• FMP inventory system collect tree species-specific data of the growing stock within stand compartments

• Stand characteristics consists of:– basal area-weighted dgM, hgM

– age (T) and basal area (G)

• Number of stems (N) is additional character, which is not required

Page 3: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

Objectives

• The objective of this study:– to examine whether the accuracy of

predicted basal-area diameter distributions (DDG) could be improved by using stem number (N) together with basal area (G)

– in terms of degree of determination (r2)– in terms of stem volume (V) and total stem

number (N), when– G is unbiased

Page 4: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

Study material• Study material consisted of:

– 91 stands of Scots pine (Pinus sylvestris L.) – 60 stands of Norway spruce (Picea abies Karst.)

• both with birch (Petula pendula Roth. and P. pubescent Ehrh.) admixtures

• in southern Finland

– about 90–120 trees/stand plot• dbh and h of all trees were measured

• Test data consisted of NFI-based permanent sample plots in southern Finland– 136 for pine– 128 for spruce– about 120 trees/cluster of three stand plots

Page 5: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

Diameter distribution

• The three-parameter Johnson’s SB distribution – bounded system includes the minimum and the

maximum endpoints – the minimum of the SB distribution () was fixed

at 0 – fitted using the ML method– to describe the basal-area diameter distribution

(DDG )

– transformed to stem frequency distribution (DDN)

Page 6: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

Distribution function

• Johnson’s SB distribution

• is based on transformation to standard normality

• in which

- z is standard normally distributed variate

and are shape parameters

and are the location and range parameters

- d is diameter observed in

a stand plot

25,0exp2

1dd zzdf

d

dz

dd

dddd

ln

ddzd

Page 7: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

Predicting the distribution

• Species-specific models for predicting the SB distribution parameters and

• Linear regression analysis

• The models were based on either – predictors that are consistent with current

FMP (ModelG)

– or those with the addition of a stem number (N) observation (ModelGN)

Page 8: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

”Percentile method”

• When predicting the SB distribution,

parameter was solved according to known and and median dgM using Formula

• Thus, known median was set for predicted distribution.

gMgM dd lnˆˆlnˆ

Page 9: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

”Shape index”

• Single stand variables: dgM, G, N or T did not

correlate closely with the shape parameter of the SB distribution

• In ModelGN, stand characteristics were linked together for ”shape index”

– in which

Ng

G

M

21004 gMM dg

Page 10: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

The behaviour of the shape index ψ

0 20 40

0.05

0.1Shape=0.94

d, cm

P

0 20 40

0.05

0.1Shape=0.89

d, cm

P

0 20 40

0.05

0.1Shape=0.76

d, cm

P

0 20 40

0.05

0.1Shape=0.74

d, cm

P

0 10 20 30

0.05

0.1Shape=0.51

d, cm

P

0 20 40

0.05

0.1Shape=0.38

d, cm

P

Stem frequency (solid line) and basal area distributions (dotted line)

Page 11: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

Correlation between parameter and shape index for spruce and pine

• Correlation r = 0.57 and 0.68 for pine and spruce, respectively

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

0 1 2 3 4 5

delta

shap

e in

dex

Spruce

Pine

Page 12: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

Results: Prediction models

• ModelG – dgM and T explained , and stem form (dgM/hgM) was

the additional variable explaining – r2 for and

• 0.22 and 0.05 for pine• 0.40 and 0.28 for spruce

• ModelGN – Shape index alone or with dgM explained and – r2 for and

• 0.28 and 0.38 for pine• 0.37 and 0.50 for spruce

Page 13: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

The relative bias and the error deviation (sbb) of the volume and stem number in the test data

ModelModelGG ModelModelGNGN

PinePine BiasBias ssbb BiasBias ssbb

V 3.0 5.1 2.4 4.9

N -4.8 12.6 -4.4 6.1

SpruceSpruce

V 1.7 6.0 2.2 5.4

N 8.7 25.0 -6.0 12.3

Page 14: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

The predicted DDGs (above) and the derived DDNs for spruce and pine, when

1.0, 0.77 and 0.63Spruce

0

0.02

0.04

0.06

0.08

0.1

0 10 20 30 40

P(g

)

0

10

20

30

40

50

60

0 10 20 30 40

d, cm

n

Pine

0

0.05

0.1

0.15

0 10 20 30 40

1.00

0.77

0.63

0

5

10

15

20

25

30

0 10 20 30 40

d, cm

Page 15: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

Advantages

• ModelGN is capable of describing great variation in N within fixed dgM and G

• Example– dgM=20 cm, G=20 m2ha-1

if = 1.00 then N = 705 and 790 ha-1

if = 0.63then N = 1020 and 1100 ha-1

for pine and spruce, respectively

Unbiased N = 640 and 1020 ha-1

Page 16: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

Height distribution

• Height distribution is not modelled for FMP purposes

• It is produced with a combination of dbh distribution and height curve models – only expected value of height is used for each dbh class

– height distribution has become of great interest lately from stand diversity point of view

• available feeding, mating and nesting sites for canopy-dwelling organisms

• Objective– to examine how the goodness of fit in marginal height

distributions can be improved using the within dbh-class height variation in models

Page 17: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

Height model including error structure

• Näslund’s height curve

• Linearized form for fitting– power =2 and 3 for pine

and spruce respectively– 0 and 1 estimated

parameters

• Residual error :– homogenous variance– normally distributed

3.1

10

d

dh

d

h

d101

3.1

Page 18: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

Error structure handling

• The residual variation (sz) of from linearized model

• transformation to concern real within-dbh-class height variation (sh)

• using Taylor’s series expansion

d

h

sszh

1

3.1ˆ

Page 19: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

Error structure behaviour

Pine

0

5

10

15

20

25

30

0 10 20 30 40

d, cm

h, m

•funtion of diameter and height

•dependent on height curve power

Spruce

0

5

10

15

20

25

30

0 10 20 30 40

d, cm

h, m

Page 20: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

Advantages

• Using expected value of h resulted in excessively narrow h variation

• Within dbh-class h variation resulted in wider h distribution

• Improved goodness of fit

• Example for pine• within dbh variation:• expected h = 22.5 to

26.0 m

• ± 2 × sh h = 19.0 to 28.5 m

0

5

10

15

20

25

30

35

0 10 20 30 40

d, cm

h, m

Page 21: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

Conclusions• Within dbh-class h variation

– reasonable behaviour with respect to dbh and h

– more realistic description of the stand structure

– improve goodness of fit of the marginal h distribution

– slight improvement with wide dbh distributions (spruce)

– significant improvement with narrow dbh distributions and strongly bending h curve (pine)

• expexted h: – 79% pass the K-S test

•including sh: –98% pass the K-S test

Page 22: Improving the accuracy of predicted diameter and height distributions Jouni Siipilehto Finnish Forest Research Institute, Vantaa E-mail: jouni.siipilehto@metla.fi

Improved accuracy and flexibility in stand structure

models

will presumably benefit modelling increasingly complex stand structures