Regional modelling of vegetation distributions in Australia · Regional modelling of vegetation...

Regional modelling of vegetation distributions in Australia Regional modelling of vegetation distributions in Australia

David HilbertCSIRO Sustainable EcosystemsTopical Forest Research CentreAtherton, Queensland

CSIRO Veg modelling Canberra 07

Prerequisites to success

• A “proper” vegetation classification for the objectives• An appropriate modelling method for the objectives• Careful data selection for training/fitting including an

appropriate range of environments and veg. types• Careful verification and awareness of limitations• Creative use of the model incorporating all ecological

knowledge

‘It s better to have the right combination of ecological expertise, local knowledge and statistical skills in the team than to be using the current “best method”’. Austin et al. 1995


Empirical modelling method

spatial distributions of vegetation classes, functional types or species

spatial distributions of environmental variables

frequency distributions of vegetation in environmental space

100

50

Probability (%)

Mean T

empe

raturePrecipitation

modelling

generalized suitability of environments for each veg. type

vegetation map


Common modelling approaches

Generalised Linear Models (GLM), GeneralisedAdditive Models (GAM), Classification andRegression Trees (CART), and Artificial NeuralNetworks (ANNs)

Uses –pre-clearing vegetationregional climate change impact assessmentpalaeoecology


2530 plots, 223 species

1. Vegetation classified into 75 “communities” 2. Statistical modelling – used GAM to predict

% cover of 135 species based on climate, topography, geology and soil variables

3. Predicted species distributions4. Used the predicted distributions and the

classification to map the vegetation communities

GAMs

Austin et al. 2000. Predicted Vegetation Cover in the Central Lachlan Region

Objective: pre-clearing vegetation map


Wet Tropics of NE Queensland Artificial Neural Network example

• Challenges – why an empirical (black box) approach?• Value of a structural vegetation classification• The Neural Net Classifier• Testing• Using all the information – various indices• Applications/further testing


The Wet Tropics Bioregion

AustraliaQueensland

c. 2 million hectares

Wet TropicsBioregion


Wet Tropics Bioregion The Wet Tropics Bioregion has:high relief (0-1600m)recent volcanismrich soilsvery high rainfallvery high biodiversity


Modelling difficulties in rainforests

• > 800 species of tree species in c. 80 families• Not all described and named• Limited and often biased biogeographic information• Very poor autecological knowledge even for the best

studied, economically important species• No useful plant functional type classification in

rainforests• Lack of predictive theoretical models• => Mechanistic approaches are impossible -

empirical approaches, focused on vegetation structure might be useful – used artificial neural network model (classifier)


“Proper” classifications of vegetation is essential

• Len Webb’s physiognomic/structural classification of RF, Specht’s classification for sclerophyll forests

• Physiognomic and structural characteristics of forests in the Wet Tropics are controlled by local environments (climate, topography, soils)

• Largely, these features are not determined by floristics

• => useful in recent past and near future (-18kyr to +50yr)


ANN Structure parametric and nonlinear

GIS

INPUTS

23 InputNodes

15 OutputNodes

Environmentalsuitability for

each of 15 forestclasses

0.9830.0000.0000.0000.0000.0030.0000.0060.0000.0030.0000.0000.0000.0000.000

HiddenLayer

(80 nodes)climategeologyterrain

Hilbert & van den Muysenberg 1999


Advantages on ANNs

• Advantages = Disadvantages

• Not analytical, no hypothesis testing, no assumptions, very good predictors with sufficient data


Data sampling is a key element Supervised Training

Training Set 75 000 pixels, 4.3% total foreststratified random areaproportional to

fA

This equalized error among vegetation types with orders of magnitude differences in area


75%at 1 hectareresolution

Acc

urac

y

Window Size (hectares)5004003002001000

0.74

0.76

0.78

0.80

0.82

0.84

Accuracy increases at

coarser scales

Modeled

Mapped

Accuracy


Apred = 10 028+ 0.913 AmappedR2 = 0.987

80060040020000

200

400

600

800

Mapped Area (103 ha)

Pred

icte

d A

rea

(103

ha)

Total areas of each forest class are predicted very well


MLW MVF

Mean annual rainfall (mm)

ANNoutput

2000 4000 6000 8000

1.0

0.8

0.6

0.4

0.2

0.0

ANN Responses

0

nonlinear, “well behaved”

SNSMVF

datarange


Additional, derived indices are useful

• Dissimilarity – index of stress• Probability of occurrence • Conversion sensitivity – considering climate change uncertainty


the index of dissimilarity is γ divided by π/2 to normalize the index to the range [0,1]

Index of Dissimilarity

ideal vector(1,0,0)

ANNoutput vector(.45,.60,.20)

γ

FT1

FT2

FT3

How “ideal” is the local environment for the forest type that is mapped there?

Hilbert & van den Muysenberg 199


Spatial estimates of the probability of finding a forest class

0

10

20

30

40

50

60

70

80

90

100

0.00 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95 1.00

Certainty Index Ranges

Prob

abili

ty o

f Tal

l Ope

n Fo

rest

(%)P(TOF) = 9.12+77.431*C(PFANN) R2 = 0.98P(TOF) = 4.87+80.735*C(FANN) R2 = 0.95

For a range of certainty values - number of pixels that are Tall Open Forest/total number of pixels

Certainty Index = ANN output TOF node x output TOF nodesum all output nodes


Applications

• Pre-clearing forest distributions• Palaeo-distributions• Landscape sensitivity – refugia• Future climate change• Linking with Cellular Automata


Pre-clearing

Cairns

Atherton

Herberton

Daintree

Mossman

Innisfail88% accuracy whentested with independent mapping of forest fragments

Hilbert (in press) 2007


Potential Today18000 BP

colddry

7000 BP

coolwet

5000 BP

warmwet

Past distributions of rainforest

Hilbert et al. 2007


SNSMHighlandRainforests

CNVFUplandRainforests

MVFLowlandRainforests

Mean stability of 9x9 window

Landscape sensitivity to climate – refugia

Inter-glacialrefugia

Glacial refugia

Hilbert et al. 2007


Future climate change

Comparing potential todaywith +1 oC, -10% rainfall

HighlandRainforests

Potential loss of the rich, endemic flora and fauna in the highlands

Hilbert et al. 2001


Lowland rainforest environments (MVF) will increase even if rainfall declines somewhat

Cha

nge

in A

rea

(%)

-80

-60

-40

-20

0

20

40

60

80

-15 -10 -5 0 5 10 15 20 25

Precipitation Change (%)

+1oC

Hilbert et al. 2001


Highland rainforest environments will decline greatly with global warming, irrespective of changes in rainfall

Cha

nge

in A

rea

(%)

-80-60-40

-200

20406080

-15 -10 -5 0 5 10 15 20 25


+1oC warming

Potentially devastating to high altitude fauna

Hilbert et al. 2004Williams & Hilbert 2006

Hilbert et al. 2001


Highland rainforests will be highly stressed by global warming, irrespective of changes in rainfall

Dis

sim

ilarit

y (in

dex

of s

tres

s)

0

0.2

0.4

0.6

0.8

-15 -10 -5 0 5 10 15 20 25

Today


+1oC warming

Hilbert et al. 2001


Spatial sensitivity to climate change Windsor Tableland

MVFMVFPSDMVFCNVFNVFSNSMDMVTVFAETOFTWMOFWMLWCCMRPAVFNSEF

0 1 2 3 4 5

+1oC warming5 rainfall scenarios

SNSM

Counts of change

boundaries &certain forest typesare most sensitive

Hilbert et al. 2001


Moving forward

Can static, empirical approaches contribute to dynamics?


Linking with Cellular AutomataCombining the output of the ANN with

spatial and ecological Information

Future environmental conditions

Future environmental suitability for individual

forest types

ANNModel

EcologicalConstraints

Neighborhood information

(updated during model iterations) Future

Vegetation

Ostendorf et al. 2001


Differences between the ANN and CA/ANN model at equilibrium for +1 oC and - 10% precipitation scenario

Spatial Constraints OnlySpatial Constraints Only Spatial and Ecological Constraints

Spatial and Ecological Constraints


Conclusions

• Regional modelling of vegetation distributions is usually empirical – often by necessity (scale issues and limited basic understanding).

• The specific modelling method is less important than the range of skills and data that are available.

• Careful empirical modelling, coupled with local ecological knowledge, good biogeographic data and creative analyses can be very informative, both scientifically and for management and policy development.

• “Black box” approaches have great value when coupled with ecological insight and are essential in the many circumstances where mechanistic models are not yet capable of addressing very important regional issues about climate change.

This research has had very significant influence on regional, state and national policy

Questions or Comments ?

David W. HilbertCSIRO Sustainable EcosystemsTropical Forest Research CentreAtherton, Queensland

Phone: +61 7 4091 8835Email: [email protected]

Contact Us Phone: 1300 363 400 or +61 3 9545 2176

Email: [email protected] Web: www.csiro.au

Regional modelling of vegetation distributions in Australia · Regional modelling of vegetation...

Documents

Transcript of Regional modelling of vegetation distributions in Australia · Regional modelling of vegetation...