Amaro A., Reed D., Soares P. (editors) Modelling Forest Systems
Подождите немного. Документ загружается.
04Amaro Forests - Chap 03 1/8/03 11:52 am Page 36
36 C. Dean et al.
estation and harvesting for sawlogs every 80 years. Stands corresponding to these
scenarios are shown in Colour Plate 4.
For the first scenario an estimated 20% of the E. regnans stem biomass, 50% of
E. regnans branches, 100% of
E. regnans leaves, 40% of the understorey biomass and
50% of decaying timber were emitted to the atmosphere during the stand replace-
ment fire. The remaining biomass was committed to the decay pools.
The reduction in soil carbon used was 32% from the data in Polglase et al. (1994)
after clearfelling of E. r
egnans. Clearfelling of wet forests such as E. regnans often
results in substantial soil disturbance, with soil inversion down to 0.3 m common on
snig tracks (log-dragging paths) and near landings (sorting, trimming, storage and
loading zones). Stand replacement fires incur less soil disturbance: we assumed a
drop in soil carbon of 10%. Typical values were used for mass extracted upon log-
ging and for its partitioning between pulp and sawlog. For example, in oldgrowth
logging 85% of the E. regnans stems were extracted and 2% of the understorey bio-
mass (with the remainder left as debris and burnt). The fraction of removed stem
pulped was 84% for oldgrowth logging and 55% for reforestation harvesting. After
the hot burn, 50% of the debris remains (Stewart and Flinn, 1985) and no live trees
remain. Typical values were used for half-lives of wood products (e.g. from Grierson
et al., 1991).
Results
The main trends following a stand replacement fire at 321 years were an initial
increase in the carbon stored in standing E. regnans, until 215 years (553 t C/ha)
(tonnes of carbon per hectare) followed by a decrease due to stand thinning and
decay. The mass from self-thinning had a peak of ~16 t C/ha/year at 215 years. The
remaining non-soil carbon from the fire (635 t C/ha) quickly decayed, causing a
rise in the humus and slow soil pool. The soil pool loses about 70 t C/ha due to the
fire then increases and levels off at 667 t C/ha at ~120 years, followed by a slight
decline to 290 years, then a slight increase due to senescing E. regnans (Fig. 3.4). The
growth of the undisturbed rainforest understorey compensates for the decay of the
ageing E. regnans and the total carbon content of the forest levels off at 1500 t C/ha
near 375 years.
The average changes (per cycle) in NBP for the stand replacement and logging
scenarios wer
e 0.80 and 2.52 t C/ha/year, respectively (with 5.90 t C/ha/year
2
after
the oldgrowth logging event). Mean values of soil carbon and total carbon for the
stand replacement fire cycle were 654 t C/ha and 1230 t C/ha, respectively. The log-
ging scenario took longer to reach dynamic equilibrium (i.e. after about five cycles),
after which its mean soil carbon and total carbon (including wood products) were
97 t C/ha and 387 t C/ha, respectively (Fig. 3.5).
Discussion and Conclusion
The modelled E. regnans stem carbon (Fig. 3.4b) falls between the values of Feller
(1980) (for a 40-year-old stand) and Goodwin (1999) (generated using a site index of
45). The curve of Grierson et al. (1991) for E. regnans biomass is slightly higher than
our curve but otherwise is similar. The data points of Feller are higher for each bio-
mass pool, possibly because we chose to be conservative in our formulation and
because his site was ideal for E. regnans.
CAR4D illustrates that carbon sequestration
04Amaro Forests - Chap 03 1/8/03 11:52 am Page 37
37 Growth Modelling for Carbon Accounting
(a) The larger carbon pools vs. stand age
(b) Stand components vs. age: a comparison
Soil carbon
standing
Fallen trees
Understorey standing
1600
1400
1200
1000
800
600
400
200
0
0 100 200 300 400
Carbon content (t C/ha)
E.r stem
Feller
E.r A-G,
Grierson
Und.,
Feller
E.r A-G
E.r stem
E.r stem, Goodwin
Und.
700
600
500
400
300
200
100
0
Carbon content (t C/ha)
Total biomass +litter +soil carbon
Total biomass + litter
Total biomass
E.regnans
Total biomass +litter
Total biomass
+litter, Feller
alone), and at 400 years
Polglase et al
stand
CAR4D shows very similar
et al
culated by
CAR4D.
half-life of the pulp component; (iv) high soil disturbance; (v)
Stand age (years)
0 50 100 150 200
Stand age (years)
et al
(a) Carbon pools vs. age
321-year stand replacement fire cycle
1600
1400
1200
1000
800
600
400
200
0
Carbon content (t C/ha)
0 200 400 600 800 1000
biomass
Understorey biomass
Slow soil + humus
(b) Carbon pools vs. age
80-year harvesting cycle
1600
1400
1200
1000
800
600
400
200
0
Carbon content (t C/ha)
0 200 400 600 800 1000
biomass
Understorey biomass
Slow soil + humus
increases beyond 100 years up to 200 years (for E. regnans
the total can approach ~1500 t C/ha, with an undisturbed rainforest type under-
storey.
. (1994) analysed soil carbon in two 53-year-old stands to about 1 m
depth and found an average of about 646 t C/ha. It increased with E. regnans
age, up to 60 years of age. The soil carbon forecast by
behaviour. Grierson . (1991) modelled an 80-year reforestation and harvesting
cycle at dynamic equilibrium but forecast slightly lower soil carbon than those cal-
Carbon pools for the harvesting cycle were low because: (i) relatively frequent
fires emit gaseous carbon; (ii) wood products were not returned to the soil; (iii) short
E. regnans were grown
only to 63% of optimum size; and (vi) the understorey was underdeveloped. The
sequestered carbon might be even lower if the following landscape scale effects
Fig. 3.4. Progress after a stand replacement fire at 321 years (without logging). Und., understorey biomass;
A-G, above ground. In (b) data from: Feller (1980), Grierson . (1991) and Goodwin (1999).
Total
E. regnans
Time (years)
Total
E. regnans
Time (years)
Fig. 3.5. Carbon pools. Starting from ten cycles of 450 years then a further 321 years of growth followed by
(a) further stand replacement fires every 321 years (without logging); and (b) clearfelling for pulp followed
by reforestation and harvesting for sawlog every 80 years. Total includes wood products.
04Amaro Forests - Chap 03 1/8/03 11:52 am Page 38
38 C. Dean et al.
were considered: escaped burns into forested areas, effect of logging roads on fire
frequency, duration between felling and germination (1 to 2 years), and the diesel
fuel used by the log skidders and loaders. These influences can be included in GIS
data for landscape-scale analyses using
CAR4D and visualized from animated output
such as in Colour Plate 5.
Acknowledgements
For assistance with data acquisition we thank: Prof D.H. Ashton (Melbourne),
Forestry Tasmania (Head and District Offices and field contractors), Parks Victoria
(Woori Yallock and Apollo Bay), Dept NRE Victoria (Research Branch), and the CRC
for Greenhouse Accounting for funding.
References
Australian Newsprint Mills (c. 1960) Unpublished coupe registers, Library of Forestry
Tasmania, Hobart, Tasmania.
Ashton, D.H. (1975a) The root and shoot development of Eucalyptus regnans F. Muell.
Australian Journal of Botany 23, 867–887.
Ashton, D.H. (1975b) Studies of litter in Eucalyptus regnans forests. Australian Journal of Botany
23, 413–433.
Ashton, D.H. (1976) The development of even-aged stands of Eucalyptus regnans F. Muell. in
Central Victoria. Australian Journal of Botany 24, 397–414.
Bassett, O.D., Edwards, L.G. and Plumpton, B.S. (2000) A comparison of four-year-old regener-
ation following six silvicultural treatments in a wet eucalypt forest in southern Tasmania.
Tasforests 12, 35–54.
Brown, B. (2001) The Valley of the Giants: a Guide to Tasmania’s Styx River Forests. Bob Brown,
Hobart, Tasmania, 66 pp.
Criminisi, A., Reid, I. and Zisserman, A. (1999) Single view metrology. In: Proceedings of
International Conference on Computer Vision, 20–25 September 1999, Corfu, Greece. IEEE
Computer Society Press, Los Alamitos, California, Vol. 1, pp. 434–441.
Feller, M.C. (1980) Biomass and nutrient distribution in two eucalypt forest ecosystems.
Australian Journal of Ecology 5, 309–333.
Galbraith, A.V. (1937) Mountain Ash (Eucalyptus regnans F. von Mueller): a General Treatise on its
Silviculture, Management, and Utilization. H.J. Green, Government Printer, Melbourne, 51 pp.
Galbraith, A.V. (1939) Proceedings of the Victorian Bush Fire Conference, August 1939.
Australian Forestry 4, 79–81.
Goodwin, A. (1999) Generalised Yield Curves for Single-aged Tasmanian Eucalypt Forest. Forestry
Tasmania, Hobart, Tasmania.
Grierson, P.F., Adams, M.A. and Attiwill, P.M. (1991) Carbon Sequestration in Victoria’s Forests
and Carbon Storage in Soil and in Forest Products. Reports commissioned by the State
Electricity Commission of Victoria.
Helms, A.D. (1945) A giant eucalypt. Australian Forestry 9, 24–28.
Hickey, J.E., Su, W., Rowe, P., Brown, M.J. and Edwards, L. (1999) Fire history of the tall wet
eucalypt forests of the Warra ecological research site, Tasmania. Australian Forestry 62(1),
66–71.
Hickey, J.E., Kostoglou, P. and Sargison, G.J. (2000) Tasmania’s tallest trees. Tasforests 12,
105–121.
Ilic, J., Boland, D., McDonald, M., Downes, G. and Blakemore, P. (2000) Woody Density Phase 1 –
State of Knowledge. Australian Greenhouse Office, National Carbon Accounting System
Technical Report No. 18.
Keith, H., Barrett, D. and Keenan, R. (2000) Review of Allometric Relationships for Estimating
Woody Biomass for New South Wales, the Australian Capital Territory, Victoria, Tasmania and
04Amaro Forests - Chap 03 1/8/03 11:52 am Page 39
39 Growth Modelling for Carbon Accounting
South Australia. Australian Greenhouse Office, National Carbon Accounting System
T
echnical Report No. 5B.
Klein Goldewijk, K., van Minnen, J.G., Kreileman, G.J.J., Vloebeld, M. and Leemans, R. (1994)
Simulating the carbon flux between the terrestrial environment and the atmosphere. In:
Alcamo, J. (ed.) IMAGE 2.0 Integrated Modelling of Global Climate Change. Kluwer
Academic, London, pp. 199–230.
Law, G. (1999) Tasmania’s Magnificent Tall Forests and the RFA. Available at:
www.greens.org.au/bobbrown/tall.htm
Mace, B. (1996) Mueller, champion of Victoria’s giant trees. The Victorian Naturalist 113(4),
199–207.
Mooney, N. and Holdsworth, M. (1991) The effects of disturbance on nesting wedge-tailed
eagles (Aquila audax fleayi) in Tasmania. Tasforests 3, 15–31.
Mount, A.B. (1964) Three studies in forest ecology. MSc thesis, University of Tasmania, Hobart,
Australia.
Olson, J.S. (1963) Energy storage and the balance of producers and decomposers in ecological
systems. Ecology 44, 322–331.
Patton, R.T. (1919) On the growth, treatment and structure of some common hardwoods.
Proceedings of the Royal Society of Tasmania 31(II), 394–411.
Polglase, P.J. and Attiwill, P.M. (1992) Nitrogen and phosphorous cycling in relation to stand
age of Eucalyptus regnans F. Muell. I. Return from plant to soil in litterfall. Plant and Soil
142, 157–166.
Polglase, P.J., Adams, M.A. and Attiwill, P.M. (1994) Measurement and Modelling of Carbon
Storage in a Chronosequence of Mountain Ash Forests: Implications for Regional and Global
Carbon Budgets. State Electricity Commission, Victoria, Australia.
Stewart, H.T.L. and Flinn, D.W. (1985) Nutrient losses from broadcast burning of eucalypt
debris in north-east Victoria, Australian Forest Research 15, 321–332.
SYSTAT (1992) SYSTAT for Windows: Statistics, Version 5 Edition. SYSTAT, Evanston, Illinois, 750
pp.
Turnbull, C.R.A. and Madden, J.L. (1986) Litter accession, accumulation and decomposition in
cool temperate forests of southern Tasmania. Australian Forest Resources 16, 145–153.
04Amaro Forests - Chap 03 1/8/03 11:52 am Page 40
05Amaro Forests - Chap 04 25/7/03 11:04 am Page 41
4 Spatial Distribution Modelling of
Forest Attributes Coupling
Remotely Sensed Imagery and GIS
Techniques
Gherardo Chirici,
1
Piermaria Corona,
2
Marco
Marchetti,
3
Fabio Maselli
4
and Lorenzo Bottai
5
Abstract
Needs for accurate information about forest resources can only partly be met by conventional
inventories based on ground sampling. Earth observation (EO) techniques are a valuable
source of information for several forest attributes which are linked to relevant spectral
responses (tree species composition, stand biomass, stand density, etc.). In particular, EO can
be effective for propagating forest inventory plot sample values through the landscape: sample
values can be assigned to non-sampled locations according to the similarity of certain spectral
features among the sampled and the non-sampled plots. A spatial modelling based on the inte-
gration of remotely sensed images and sample field measurements targeted to produce forest
attributes maps is presented for a site in central Italy with more than 300 geocoded sampling
field plots. Plot data of tree stemwood volume and other non-wood forest attributes came from
a single-stage cluster design with 58 primary sampling units (clusters). Landsat 7 ETM+ images
are used with two classification techniques (k-NN and fuzzy classifiers) to model the spatial
distribution of stemwood volume (m
3
/ha) and stem density (n/ha). Modelling and mapping
results are discussed.
Introduction
Forest monitoring programmes are recognized as a main information support for
terrestrial natural renewable resource modelling and management. Distinctively, the
sustainable management of forest resources requires a large amount of supporting
information. Traditional ground-sampling-based forest inventories provide sum-
mary statistics of forest attributes. However, given the usual sampling intensities,
inventory information cannot directly support an estimation and visualization of
the measured attributes (e.g. tree species composition, stand density, stand height,
1
geoLAB, Dipartimento di Scienze e Tecnologie Ambientali Forestali, Università di Firenze, Italy
Correspondence to: gherardo.chirici@unifi.it
2
Dipartimento di Scienze dell’Ambiente Forestale e delle sue Risorse, Università della Tuscia, Italy
3
Dipartimento di Colture Arboree, Università di Palermo, Italy
4
IATA, Consiglio Nazionale delle Ricerche, Italy
5
LAMMA, Regione Toscana, Italy
© CAB International 2003. Modelling Forest Systems (eds A. Amaro, D. Reed and P. Soares) 41
05Amaro Forests - Chap 04 25/7/03 11:04 am Page 42
42 G. Chirici et al.
stand volume, etc.) at the local level. A complete mapping of forest variables and
associated characteristics would greatly benefit forest management planning. A
complete mapping of forest attributes also provides a basic information source for
many modelling tasks such as animal habitat suitability assessment, hydrological
applications, and so on.
Earth observation (EO) techniques at various scales (Barbati et al.
, 2000) play a
distinctive role in the effective development of environmental mapping pro-
grammes. In the specific sector of natural resource surveys, it is important to evalu-
ate change. Consequently, frequent coverage is essential, and satellite remote
sensing can provide data which are cost-effective and not available from any other
source. In particular, EO can be effective for propagating forest plot inventory val-
ues through the landscape; sampled values can be assigned to the remaining non-
selected plot locations according to their spectral features (Franco-Lopez et al., 2001).
Non-parametric classification methods are intrinsically suited to this objective, since
they allow a flexible treatment of the remotely sensed data without assuming pre-
determined statistical distributions which are unrealistic in many forest environ-
ments (Richards, 1993; Fazakas et al., 1999).
The present research aims at assessing the potential of known non-parametric
classification methods to model the spatial distribution of for
est attributes relying
on Landsat 7
ETM+ data. In particular, the investigation was carried out in a forested
area in central Italy for which a routine ground forest inventory had recently pro-
duced more than 300 sampling points grouped in 58 clusters. Inventory data were
used for both calibrating and verifying the performances of two EO classification
methods, based respectively on k-NN and fuzzy classifiers.
Experimental Test Site
The test site is located in the Municipality of Acquapendente, within the
Mediterranean biogeographical region in central Italy. The site is about 8600 ha, and
completely includes the Riserva Naturale di Monte Rufeno, a regional designated
conservation area of about 3000 ha. The site has been chosen both for its representa-
tiveness of Mediterranean hilly conditions and for the availability of recent studies
regarding several environmental aspects. The area is characterized by a fairly
smooth morphology with slopes ranging around 30–35%.
The most common vegetation types are: forests (oak woods with Quercus cerris
pr
evailing), shrubs (mantle communities and bushes), grassland, pasture and
garigue, riparian shrub and tree vegetation (with alder, poplar and willow), agricul-
tural fields and forest plantations (various species of fast-growing conifers). Forests
within the Riserva Naturale di Monte Rufeno are in conversion to high forest with
the application of selective thinning cuts. Most coppices outside the Riserva are still
managed for the production of fuelwood, with a 15-year rotation. The most com-
mon silviculture system is coppice with standards. Some coppices are abandoned
and revert to their natural evolution. Conifer plantations were established between
1950 and 1970, using various species: including Pinus pinea, Pinus excelsa, Pinus nigra
and Cupressus arizonica. Some of the plantations have been thinned, whereas others
still have a very high density. The Mediterranean macchia is well represented in the
area.
For the surveyed forestland, the average forest coverage is about 65%, the aver-
age stand density is close to 3500 stems/ha, the average basal ar
ea is 18 m
2
/ha and
the average stand volume is 128 m
3
/ha.
05Amaro Forests - Chap 04 25/7/03 11:04 am Page 43
43 Spatial Distribution Modelling
Data
Ground sampling data
Ground sampling was configured as a single-stage systematic cluster design. On the
basis of the UTM geographical r
eference system, a kilometric sample grid was gen-
erated to locate plot clusters. Two different kind of clusters were displaced on the
grid: ‘dense’ clusters, made of nine sample plots in a 3 × 3 configuration, and
‘sparse’ clusters, made of five sample plots in an L-shaped configuration. Each sam-
ple plot has a circular shape, with a 10-m radius. In dense clusters, sample plots are
25 m apart, while in sparse clusters they are 100 m apart (Fig. 4.1).
A first selection of forested clusters was carried out on the basis of a photo-
interpr
etation of digital aerial orthophotographs from 1996. Photographs were ren-
dered in an 8-bit greyscale with a spatial resolution of 1 m. After this selection, 31
cluster sampling points (i.e. the origin of each cluster) were classified as non-
forested by photo-interpretation and excluded from further analysis. Only the
remaining 58 clusters (38 sparse and 20 dense) were used for analyses (Fig. 4.2).
Satellite data
Satellite data consisted of a Landsat 7
ETM+ scene taken on 31st July 2000. The scene
was orthocorrected using more than 50 ground control points and a 75-m resolution
digital elevation model. The global positional accuracy was less than the nominal
side length of a pixel. Subsequently the scene was superimposed on the ground
sampling grid which reported the 370 sample plots. No topographic normalization
was performed because the spectral effect of topographic irregularities was consid-
ered to be negligible in the gently rugged terrain.
Spectral signatures of all forest sample plots were extracted from
ETM+ bands 3,
4 and 5 as recommended by Horler and Ahern (1986).
Fig. 4.1. Configuration of the sparse (bottom left) and dense (upper right) primary sampling units (clusters).
05Amaro Forests - Chap 04 25/7/03 11:04 am Page 44
44 G. Chirici et al.
Fig. 4.2. Geographical layout of single-stage systematic cluster sampling carried out by the forest inventory
(Acquapendente, central Italy). Cluster locations are displayed on the mosaic of digital orthophotographs:
triangular dots show the locations with the sparse configuration of sample plots, while square dots locate the
dense configuration of sample plots. Boundaries of the study site are shown by the wider white line, while
boundaries of the Riserva Naturale Regionale di Monte Rufeno are indicated by the thinner white line.
Methods
Tested classifiers
The spectral signatures of all sample plots were used within two popular non-
parametric classification techniques (Fazakas et al.
, 1999; Katila and Tomppo, 2001;
Maselli, 2001; Maselli et al., 2001). Both classifiers are flexible, and allow a series of
05Amaro Forests - Chap 04 25/7/03 11:04 am Page 45
45 Spatial Distribution Modelling
options which can adapt their application to the specific environmental situation
and data used.
The first classifier is a k-nearest neighbour (k
-NN) procedure, which has been
used by Finnish and Swedish forest experts for updating their national inventories
(Fazakas et al., 1999; Katila and Tomppo, 2001). In the k-NN estimation procedure,
the variables for a specified pixel are predicted as weighted averages of the most
spectrally similar reference pixels (the k nearest neighbours). Several options are
allowed regarding the form of spectral distance computed and the thresholding of
spatial and altitudinal distances (Franco-Lopez et al., 2001; Katila and Tomppo,
2001). Our version of k-NN used the spectral Euclidean distance as the discriminat-
ing feature, with the number k determined in each single case (see below). Due to
the limited size and topographic variability of the study area, no horizontal or verti-
cal thresholding was applied.
The second classifier is a simplified version of the fuzzy discrimination proce-
dur
e recently proposed by Maselli (2001). This uses the same Euclidean distance as
before to compute for each estimation point spectral weights with respect to all ref-
erence pixels. These spectral weights W are defined as:
s
(E
d
/R
d
)
(1)W
s
= exp
where E
d
is the spectral Euclidean distance between the estimation and measured
points, and R
d
is a constant (range) which regulates the weighting of the spectral dis-
tances. A very large R
d
effectively assigns a weight of 1 to all observations, while a
small R
d
assigns a weight that decreases rapidly with the spectral distance. The main
difference between the two estimation methods is the inclusion of all reference points
in the fuzzy classifier (even though a spectral threshold can also be applied). As can
easily be understood, this approach is conceptually similar to kriging, which uses
estimates of the covariance between the location to be estimated and all sampled
locations to compute a prediction that is a weighted averages of all reference points
for each point examined (Davis, 1973). Similarly to kriging, the determination of the
spectral range is a crucial step in the procedure, which was currently carried out for
each parameter as described below.
Forest attributes and classifiers configuration
The modelling experimentation focused on the identification of the best configura-
tion of the two classifiers for tr
ee stemwood volume (m
3/
ha) and tree stem density
(n/ha). The fine-tuning of the classifers was done by a leave-one-out strategy (Franco-
Lopez et al., 2001), which allowed the consideration of almost all reference pixels: it
must be recalled that both classifiers, being intrinsically non-parametric, are very sen-
sitive to the representativeness of the training sample, which must therefore be as
large as possible for a correct accuracy assessment. Application of the leave-one-out
strategy allowed the identification of the k and R
d
values that minimize the average
prediction error for a single location.
Results
k-NN classifier
The results of the leave-one-out application of the
k-NN classifier are summarized in
Fig. 4.3. As can be seen, attractive configurations of the classifier are found with