You are on page 1of 21

FUEL HAZARD MAPPING OF THE VICTORIAN CENTRAL

HIGHLANDS USING LIDAR DATA


1 1 2
Andrew Haywood , Andrew Mellor , Anders Siggins

1
Victorian Department of Sustainability and Environment
PO Box 500, East Melbourne, Victoria 3002
Phone 03 9637 9680, Fax 03 9637 8117
Andrew.Haywood@dse.vic.gov.au
2
CSIRO Sustainable Ecosystems
Box 312, Clayton Victoria 3169

Abstract
Airborne LiDAR data procured by DSE Victoria in 2007 and 2008 over the
Victorian Central Forest Management Area and regions of the Murrindindi Shire
has been processed for fuel hazard mapping. The LiDAR data covers an area
of approximately 480 000 hectares and includes significant areas of forest burnt
in the bushfires of February 2009. The data was processed using the CSIRO
LiDAR ‘ToolBox’ which produced a range of forest structural layers in a 10 m
grid. Output layers included indicators of vertical structure based on vertical
cover by height strata and skewness in the distribution of intercepted LiDAR
hits within the canopy.

A Principal Component Analysis (PCA) was conducted on the output layers.


The PCA was performed separately on the burnt and unburnt areas within the
extent of the LiDAR data. The burnt and unburnt PCA ‘signatures’ were
compared and it was decided that they were similar enough that fuel hazard
sampling in unburnt areas in the summer of 09/10 could be used to explain pre-
fire fuel loads in burnt areas.

The results of this project show the utility of LiDAR and Landsat TM data in
retrospective fuel hazard mapping.

1
Introduction
LIDAR (Light Detection and Ranging) systems utilize lasers and detectors in
various configurations to make accurate measurements of platform to surface
distances. LIDAR techniques have been used within Australia on a variety of
platforms to estimate forest height and canopy structure (e.g., Lee et al., 2004;
Turner, 2007). Specific applications have been to produce forest inventories
(Tickle et al. 2001; Turner and Webster, 2005, Turner 2006, Rombouts et al.,
2008; Haywood et al. 2009; Rombouts, 2009). Recently, LIDAR analysis
techniques have been used to estimate fuel load/hazard parameters for
Australian forests (Roff et al. 2006, Sutherland et al., 2006).
LIDAR has the potential to be an important tool for fire managers in forested
regions. However, one of the problems in using LiDAR data and their
associated products is that they are complex, expensive and their processing is
often beyond the technical expertise of most land management agencies.
Adopting remote sensing technologies in the management process requires
investment in hardware, software and training. Before a group of users adopt
this technology as an alternative to existing methods they need to be convinced
that the remote sensing solutions are economically viable and save time and
money or enable the realisation of environmental benefits.
The purpose of this study was to demonstrate the benefits of using LiDAR data
in fuel hazard applications, particularly in relation to cost and time resources.
The study has focussed on the following principles to aid the adoption of this
technology within the Land and Fire Division of the Victorian Department of
Sustainability and Environment.
• Leverage of existing imagery datasets – due to financial and budget
constraints this study has utilised existing LiDAR data capture from the
summer of 2007/08.
• Early and continuous involvement of end users – from the outset the
project team has involved end users, which is critical for the success of
the application development and ultimately for realising an impact on
decision making.
• Provide training and education – at present LiDAR data/technologies
are not widely used within the Land and Fire Division of the Victorian
Department of Sustainability and Environment, and training and
educating these users is critical for the adoption of these technologies.
After the development of a fuel hazard model, the routine operational
activities must be handed over to the end-users or some other
intermediary, thereby freeing up the research community to pursue
additional investigations. The basic model adopted within this study was
that the scientist would develop and prototype the application by working
with the end users. After successful demonstration of the benefits of
LiDAR derived products, these products and methods would be
transferred to the relevant end-users who would then adopt and maintain
them. Training and education are, therefore, critical to transfer of

2
technology from the application developers to the end users with the
land management agencies.
• Modelling Approach – A straightforward approach to the analysis was
applied using open-source software and well established techniques for
data preparation and classification.
• Utilisation of standard formats and freely accessible products –
remote sensing data and products are generated by various land
management agencies in a variety of formats. These formats may be
suitable for researcher, but end users in non-research organisations
require data in a format that can be readily displayed and analysed using
simple software packages. End users within land management agencies
typically do not have access to high-end image processing software and
tools. Products from the LiDAR application developed within this study
should, therefore, be generated in formats that can be readily ingested
into a variety of popular low cost image visualisation and GIS packages
and should be compatible with user’s existing systems and
infrastructure.
For the purposes of this study we define the “overall fuel hazard” as the sum of
the influences of bark hazard, elevated fuel hazard and surface fine fuel hazard
(McCarthy et al. 1999). This approach represents a deviation from the generally
held philosophy of assessing fuel factors by simply considering the surface fine
fuels. The “overall fuel hazard” shifts the emphasis to considering the whole
fuel complex, and particularly the bark and elevated fuels – which are
considered the fuel elements principally responsible for both first attack failure
and also for general suppression difficulty in Victorian forests, woodlands,
deserts, heathlands and shrublands (McCarthy and Tolhurst, 1998).

The overall aim of this study is to develop a methodology to use LiDAR data
and ground based assessments to accurately and efficiently map “overall fuel
hazards” in the Central Highlands in Victoria. The specific objectives of this
study are as follows:
1. Develop LiDAR-derived products and the methodology to use them for
modelling fuel hazards across large areas;
2. Investigate the use of data fusion techniques for modelling fuel hazards;
and
3. Produce spatially explicit digital fuel maps.

3
Methods

Study area
Field inventory plots were established in unburnt forests in the Central
Highlands in Victoria (Figure 1). The study area covers a variety of forest types,
with Ash species (mountain ash (Eucalyptus regnans) and alpine ash
(Eucalyptus delegatensis)) occurring in the higher elevations. As the elevation
decreases, these forest structure species are replaced by messmate
(Eucalyptus obliqua), and in drier areas, Peppermints, Boxes and Candlebark.
In general, the area experiences a cool temperate climate, with mild summers
and cool winters. Average annual rainfall exceeds 1200 mm over most of the
area. Soils tend to be free draining, friable, brown gradational, have high water
holding capacities, and have developed on a variety of volcanic parent rock
materials (DNRE 1998).
The study area is subject to intensive hardwood timber harvesting. Large-scale
timber cutting, generally selective harvesting and sawmilling occurred in these
forests in the latter part of the nineteenth and early twentieth centuries. Massive
salvage operations have followed major wildfires. Since the 1960s, clear-felling
has been the major silvicultural system practised (Squire et al. 1991).
Fire is the major natural disturbance associated with the study area. Several
fires have occurred over the past 150 years, the most extensive being in 1926,
1939 (Jeremiah and Roob 1992) and the recent fire event of February 2009.

Figure 1: The location of the Central Highlands Area, Victoria Australia, including the
extent of the LiDAR coverage, public land forests and the Kilmore East-Murrindindi
complex area burnt during the 2009 February fires.

4
Sampling Design and Field Data
In order to assess the fuel hazard models and forest inventory parameters and
determine the accuracy of airborne LiDAR estimate, ground inventory data
were gathered for this study between April and June 2010. Sampling was
based on a stratified random sample with strata defined using a PCA. The
sample population was initially reduced by excluding areas with slopes greater
than 25° and more than 200 m from a forest road or trail. The first PCA of the
LiDAR stack (refer Image Processing Section) was extracted and an
unsupervised classification performed to produce 10 spectrally homogeneous
strata classes. Candidate sampling locations were further reduced by selecting
strata areas with a minimum size of 75 pixels (3 ha). Stratified random sampling
was applied to allocate 65 plots to the 13 strata (with an extra 35 plots for
purposive sampling, i.e. easy to access areas – for a total of 100 plots). Sixty
five plot coordinates were then generated for field assessment and a set of
additional polygons for the purposive sampling. It should be noted that plots
located on escarpments by this process were relocated due to safety concerns
for the field personal. At the time of publication only 42 (out of the 100 total)
had been sampled. The analysis shown in this paper is therefore only an
interim result and will be re-run when the field sampling is complete.
The centre of each plot was located using a standard global positioning system
and pegged with flagging tape. Starting from a central point (plot location), three
50 m transects were established at 45° (North East), 180° (South) and 315°
(North West). A fuel hazard assessment was undertaken within a 10 m radius
of the end of each 50 m transect arm (sub-plot), to derive average ratings
(scores) – Low (1), Medium (2), High (3), Very High (4), Extreme (5) – for
surface hazard, near surface hazard, elevated fuel hazard, bark hazard, from
which is derived an overall field hazard score and rating. Appendix 1 includes
descriptions of each fuel layer and the fuel assessment field data sheet, which
describes the system for rating (scoring) each of the fuel layers.

5
LiDAR and Landsat Data
Moderate-density LiDAR data were acquired over the study area with an
Optech ALTM 3100 EA system (AAM Hatch Ltd) during December 2007 and
January 2008. The system settings and flight parameters are shown in Table 1.
The vendor provided raw LiDAR data consisting of XYZ coordinates, off-nadir
scan angle, and intensity for all laser returns within the area in LAS file format.
In addition, the vendor provided a data set consisting of the subset of laser
points presumed to be measurements of the terrain surface, identified via a
proprietary filtering algorithm. These filtered ground returns were used to
generate a 1 m digital terrain model (DTM). This resolution was selected for
landscape hydrological purposes beyond the scope of this study and it is
acknowledged it may not be appropriate for much of the vegetated study area.

Table 1: LiDAR data specification

Attribute Value
Flying Height 1 300 m
Swath Width 945 m
Swath Overlap 25%
Flight Direction Ne/SW
2
Laser Pulse Density 0.96 pulses/m
Laser Footprint Size 0.26 m

A level 1T (L1T) processed Landsat TM scene (WRS 092086) from January


2008, covering the extent of the LiDAR region, was acquired from USGS Earth
Observation and Science Center (EROS).

6
Data Processing
Figure 2 shows the data processing and modelling undertaken to derive fuel
hazard maps of the study area.

Field Data LiDAR (raw) Landsat TM (raw)

Data Processing Data Processing


Data Processing (LiDAR toolbox) (GRASS)

Dependent variables Independent variables

Fuel Hazard Score LiDAR metrics Landsat TM metrics

Data Fusion Approach

1. LiDAR-Landsat Stack 2. Principal Components Analysis

Principle Components % Cover (Lidar)


% Cover (Lidar) Supervised Image Classification Lidar-NDVI-PCA.1 • 0 - 0.5 m
• 0 - 0.5 m (Gaussian maximum likelihood Lidar-NDVI-PCA.2 • 0.5 - 1.0 m
• 0.5 - 1.0 m classification – GRASS) Lidar-NDVI-PCA.3 • 1.0 - 1.5 m
• 1.0 - 1.5 m Lidar-NDVI-PCA.4 • 1.5 - 2.0 m
• 1.5 - 2.0 m Lidar-NDVI-PCA.5 • 2.0 - 2.5 m
• 2.0 - 2.5 m • 2.5 - 3.0 m
• 2.5 - 3.0 m Majority filter (GRASS) • 3.0 - 4.0 m
• 3.0 - 4.0 m • 4.0 - 5.0 m
• 4.0 - 5.0 m • 95th percent. height
• 95th percent. height Fuel Hazard Prediction Map • Skewness
• Skewness Landsat TM metric
Landsat TM metric • Landsat NDVI
• Landsat NDVI
Accuracy Assessment Cross-validation

Figure 2: Flow chart of processing approach


Ten LiDAR surfaces were generated using the CSIRO LiDAR Toolbox and
used as inputs to the fuel hazard prediction model. LiDAR surfaces comprised
10x10 m cells, each attributed a percentage cover at the represented height
strata – calculated from the number of points within the 10x10 m cell within the
height strata (e.g. 50cm to 1 m) divided by the total number of points that fell
within the cell. Surface strata applied were 0-0.5 m, 0.5-1.0 m, 1.0-1.5 m, 1.5-
2.0m, 2.0-2.5 m, 2.5-3.0 m, 2.5-3.0 m and 3.0-4.0 m. Percentage cover in the
95th height percentile and skewness across all height bins were also used in the
analysis.
Landsat TM bands 3 and 4 were resampled to 10 m resolution and used to
derive a Normalized Difference Vegetation Index (NDVI).

7
Data Fusion Approach
Data fusion deals with association, correlation, and combination of information
and data from one or many sources (Llinas, 2002). In this study, two
approaches to fusing data were applied and used to model fuel hazard: A
LiDAR-multispectral (NDVI) stack and a Principal Component Analysis (PCA).
The LiDAR-multispectral stack was created by stacking the LiDAR cover, height
and skewness surfaces together with the NDVI surface. PCA is a helpful
statistical technique that is used to produce uncorrelated output bands, to
segregate noise components, and to reduce the dimensionality of data sets
(Jensen, 2005). We investigated these two different approaches to create new
fused images. We then applied supervised image classification on the new
images, examined its effect for improving overall classification accuracy, and
finally compared results.

LiDAR-Landsat derived stack


By using Grass GIS (http://grass.itc.it) we built a new multiband image with 10
m spatial resolution. This image includes a total of 11 bands and will be
subsequently referred to as the LiDAR-Landsat Stack. We used only the first
eight LiDAR bins by assuming they characterize best the vertical structure of
elevated fuels within a 5 m vertical canopy space adjacent to the ground. The
LiDAR 95th height percent and skewness coupled with an NDVI Landsat tm
index were also included.
Principal Component Analysis (PCA)
PCA was applied to the ten band LiDAR-Landsat stack image. We used the
first five of the ten PCs for our subsequent image classification. The PCA
transformation is based on the variance and covariance of the data set (Chavez
et al., 1991). Eigenvalues, variance, and eigenvector were extracted for each
PC. Table 2 represents the percentage of total variance, eigenvalues, and
cumulative variance explained by each principal component. The first principal
component accounts for 74.8% of variance in the entire LiDAR and
multispectral data set. In addition, the first 5 components that we used for
image classification account for approximately 98.5% of the total variance. It
can be concluded that the first five principal components can replace the
original eleven bands of the LiDAR-Landsat stack image, while reducing the
size of the data set, redundancy, and noise.
Table 2: Calculations of eigenvalues, percentage of total variance and cumulative
percentages for each Principal Component

PCs Eigenvalue % of total Cumulative


variance
1 615.2 74.8% 74.8%
2 131.9 16.0% 90.8%
3 40.6 4.9% 95.8%

8
4 15.8 1.9% 97.7%
5 6.2 0.8% 98.5%
6 5.3 0.6% 99.1%
7 3.4 0.4% 99.5%
8 1.6 0.2% 99.7%
9 1.1 0.1% 99.8%
10 0.8 0.1% 99.9%
11 0.5 0.1% 100.0%

Burn versus Unburnt Analysis


Before any field work was completed a Principal Component Analysis (PCA)
was conducted on the LiDAR-Landsat stack. The PCA was performed
separately on the burn and unburnt areas within the extent of the LiDAR data.
The burnt and unburnt PCA ‘signatures’ were compared and it was decided that
they were similar enough that fuel hazard sampling in unburnt areas in the
summer of 09/10 could be used to explain pre-fire fuel loads in burnt areas.

Image Processing
The first step in undertaking a supervised classification is to define the areas
that will be used as training sites for each of the fuel hazard classes. The region
of Interest (ROI) actually corresponds to our field plots. A plot of 80 m radius
covers approximately 201 pixels, given a 10 m spatial resolution for our data.
Thus the 201 pixel size resulted from collecting a circular shaped ROI over the
field plot.
A total of 42 polygons were selected for the image classification training set.
The training set results in a total of 8442 (42 * 201) pixels for both the LiDAR-
Landsat stack image and the principal component image. ROI separability
reports, which calculate the spectral separability between chosen ROI pairs for
a given input file (Jensen, 2005) were computed for each dataset. The same
ROIs chosen for the LiDAR-Landsat stack image were used on the stack of first
five PC’s.
The average overall fuel hazard score of each training plot was calculated from
the three sub-plots to be used as ground training data value. Each circular plot
polygon was given a fuel hazard score attribute based on the position of the
average within the hazard score ranges: 2 (Medium); 3 (High); 4 (Very High); 5
(Extreme).
Supervised image classification was performed using the parametric decision
rule Maximum Likelihood with each of the image stacks. According to Jensen
(2005, p. 350), the training data set is of value if the environment from which
they were collected is relatively homogenous. As such, 3 field training plots

9
were not used for either training or the accuracy validation datasets as these
plots were located in non-homogenous forest stands. Since we evaluated fuel
hazard on a per plot basis, we applied a majority filter to the classified image
using a 3 x 3 window size, to account for positional errors in GPS sub plot
centre points.
We assessed the accuracy of the fuel hazard map for each data set using
cross validation. Although data splitting has been proposed as a way of
validating models (Snee, 1977; Miller, 1984; Picard and Cook, 1984; Picard and
Berk, 1990), it has been demonstrated to be a costly approach, both in terms of
predicted accuracy of the fitted model and the precision of the estimates of that
accuracy (Roecker, 1991). Therefore, the accuracy of the model
parameterisations was assessed using the ‘leave-one-out’ cross-validation,
which is a standard tool for estimating prediction error (Stone, 1974; Allen,
1974). This process involved excluding an individual plot from the training data,
classifying the remaining training plots and comparing fuel hazard predictions
using the resulting parameter estimates with actual data from the excluded
training plot. This procedure was performed for each training plot in turn, and
prediction errors were examined to measure the precision of each
parameterisation/region. This was done for both prediction maps. The Kappa
statistic was used for the resultant accuracy assessment. The Kappa statistic
(K-hat) derived for each classification verifies if classification results are precise
(Jensen, 2005).

Results
The results of the two classification methods (and associated accuracies) were
assessed.

LiDAR derived stack


Table 3 represents the confusion matrix for the LiDAR-Landsat stack image
classification.
Table 3: Confusion Matrix for the LiDAR-Landsat stack image classification.

Ground Reference Class


2 3 4 5 Total
2 2 3 1 0 6
Mapped 3 1 2 3 1 7
Class 4 0 2 5 0 7
5 0 3 6 10 19
Total 3 10 15 11 39

10
Table 4 shows per-class user’s accuracies and producer reliabilities along with
their standard deviations and confidence intervals (both approximate and
exact). Note that the exact confidence intervals differ substantially from the
approximation in the classes with small sample size.
Table 4: Naïve per class statistics for the LiDAR-Landsat stack image Confusion Matrix

Class User’s Accuracy

Ci si 95% C.I. Exact C.I.

2 0.3333 0.1925 0.0432…0.7772

3 0.2857 0.1707 0.0366…0.7095

4 0.7143 0.1707 0.2904…0.9633

5 0.5263 0.1145 0.2886…0.7555

Class Producer’s reliability

Ci si 95% C.I. Exact C.I.

2 0.6667 0.2722 0.0942…0.9915

3 0.2000 0.1265 0.0252…0.5560

4 0.3333 0.1217 0.1182…0.6161

5 0.9091 0.0867 0.5872…0.9977

Overall Accuracy: 0.4872


s= 0.08; 95% C.I.: 0.3175 ... 0.6569

We can examine the confusion matrix and statistics to gain insight on the
magnitude and type of errors. Clearly, some hazard classes are better mapped
than others. From the map user’s point of view, the classes are mapped with a
range of accuracy between 30% to 70%. From the map producer’s point of
view, only classes 2 and 5 were reliably mapped.

11
The Kappa coefficients and statistics from the confusion matrix in Table 3 are
shown in Table 5.
Table 5: Kappa statistics for the LiDAR-Landsat stack image Confusion Matrix

Statistic Value

Number of observations 39

k̂ 0.3199

s[ k̂ ] 0.0523

C.V 16.4

95% C.I. 0.2143… 0.4256

In this case, k̂ (≈ 32%) is substantially lower than the naive measurement of


agreement, namely overall accuracy (≈ 48%), indicating that a large portion of
the apparent classification accuracy could be due to chance agreement.
Because of the small sample size, the confidence interval for k̂ is wide. This is
also shown by the high coefficient of variability. These are decreased as the
square root of the increase n sample size. For example to double the precision,
the sample size must be quadrupled.

Figure 3 shows the results of multispectral LiDAR–Landsat TM NDVI stack


image and LiDAR–Landsat TM NDVI PCA classifications

12
Figure 3: The fuel hazard classification results of LiDAR-Landsat NDVI PCA (top) and
stack (below)

13
PCA derived stack
Table 6 represents the confusion matrix for the PCA stack image classification.
Table 6: Confusion Matrix for the PCA stack image classification.

Ground Reference Class


2 3 4 5 Total
2 2 2 0 0 4
Mapped 3 1 4 3 0 8
Class 4 0 2 5 0 7
5 0 2 7 11 20
Total 3 10 15 11 39

Table 7 shows per-class user’s accuracies and producer reliabilities along with
their standard deviations and confidence intervals (both approximate and
exact). Note that the exact confidence intervals differ substantially from the
approximate in the classes with small sample size.
Table 7: Naïve per class statistics for the PCA stack image Confusion Matrix

Class User’s Accuracy

Ci si 95% C.I. Exact C.I.

2 0.5000 0.2500 0.0675… 0.9324

3 0.5000 0.1768 0.1570… 0.8429

4 0.7143 0.1707 0.2904… 0.9633

5 0.5500 0.1112 0.3152… 0.7694

Class Producer’s reliability

Ci si 95% C.I. Exact C.I.

2 0.6667 0.2722 0.0942… 0.9915

3 0.4000 0.1549 0.1215… 0.7376

4 0.3333 0.1217 0.1182… 0.6161

5 1.0000 - -

Overall Accuracy: 0.5641


s= 0.0794; 95% C.I.: 0.3957 ... 0.7327

14
We can examine the confusion matrix and statistics to gain insight on the
magnitude and type of errors. Clearly, some hazard classes are better mapped
than others. From the map user’s point of view, the classes are mapped with a
range of accuracy between 50% to 70%. From the map producer’s point of
view, only classes 2 and 5 were reliably mapped.

The Kappa coefficients and statistics from the confusion matrix in Table 6 are
shown in Table 8.
Table 8: Kappa statistics for the PCA stack image Confusion Matrix

Statistic Value

Number of observations 39

k̂ 0.3995

s[ k̂ ] 0.0999

C.V 25

95% C.I. 0.1909 ... 0.608

Once again, k̂ (≈ 40%) is substantially lower than the naive measurement of


agreement, namely overall accuracy (≈ 56%), indicating that a large portion of
the apparent classification accuracy could be due to chance agreement.

Discussion
The results of this study indicate that LiDAR and Landsat can be used to
generate moderately accurate estimates of fuel hazards. It appears that the
LiDAR based forest metrics, based on the height distributions of LiDAR
measurements, capture structural information related to quantitative fuel hazard
characteristics. Results of the cross-validation procedures, as well as qualitative
assessment of the fuel hazard maps, indicate that the models have reasonable
predictive value over the extent of the study area.
It appears that the PCA fusion method provides a more accurate approach,
with an unweighted kappa index of agreement of 40% (versus 32%) and a
naïve measurement of agreement of 56% (versus 49%).
There are a number of possible sources for the discrepancy between the
remote sensing based fuel hazard metric within a plot area and the field based
metric. First, surface fine fuels, bark and elevated fuels are not directly
measured in the field but are “visually assessed”. This could introduce a
significant source of variability into the field based fuel estimates. Differing
supporting areas could also lead to significant differences between the remote
sensing and field based estimates. The field based estimates are from 3 sub-

15
plots of 10 m radius, with a total area of 0.03 hectares. In contrast, the remote
sensing training areas are 80 m radius plots covering a broader area of forest
vegetation (2.0 ha). These areas may represent different forest structure which
may increase the variability.
Another possible reason for a discrepancy between the remote sensing and
field based estimates is the nature of LiDAR data. LiDAR data represents all
vegetation components, including foliage, stems and branches. Furthermore,
the relative frequency of stems and large branches measurements increases
with a lower stem density, since more laser pulses are able to penetrate
through canopy openings. However, in the context of hazard fuel analysis,
stems and large branches are not considered fuel hazards. This may lead to a
negative bias in the remote sensing estimate of fuel hazard in less dense
stands when compared to the field based estimates.
When implementing the remote sensing supervised classification approach to
mapping fuel hazard variables as described here, it is critical to acquire field
data over the full range of forest stand types present in the area to be mapped.
Estimates of fuel hazard variables in areas with different stand structures from
those sampled are extrapolations outside the domain of the field data and are
unreliable. In this study an attempt to avoid this was done by allowing the
spectral and structural signatures of the LiDAR and Landsat data to guide and
direct the ground based data collection efforts (sampling frame). The intent
here was to avoid invalid sites and focus on those sites necessary to build a
detailed ground based data set that accurately represents the forest stand
types over the entire project area. Unfortunately, only 40% of the ground data
collection was completed at the time of publication and as such not all forest
stand types have been represented (only 4 out of potential 5 fuel hazard
classes were present). The results from this study will be updated once all
ground plots have been collected. As such these results should be viewed as
interim.
It should also be noted that the variability in the forest stand structures present
within the Central Highlands study area is not necessarily representative of the
natural structural variability with Victoria’s forests. Therefore, the classification
models developed in this study are meant to demonstrate the potential of this
methodology for fuel hazard estimation, and do not necessarily reflect
fundamental physical relationships between LiDAR/spectral distributions and
biophysical properties of natural stands.

Conclusion
The results of this study indicate that LiDAR fused with Landsat can be used to
generate estimates of fuel hazards efficiently and accurately over an extensive
area within the Central Highlands of Victoria. Fuel hazard estimates based
upon the distribution of LiDAR measurements and multispectral Landsat data
can be used to generate maps that produce a spatially explicit distribution of
fuel hazard over the landscape. These maps can serve as a direct input into
fire-behaviour models, potentially enabling a more realistic and accurate
prediction of fire spread and intensity.

16
In the future, this methodology could be applied to LiDAR and multispectral
data collected in different forest stand types, including high risk areas in
Victoria. A more rigorous model validation procedure should be carried out to
assess the general applicability of these models in different forests types. It is
likely that a more extensive pool of explanatory variables will be developed to
improve our understanding of the structural relationship between the distribution
of LiDAR measurements and multispectral data with fuel hazard
measurements.

17
Acknowledgements
This study was supported by the Victorian Government. We would like to thank
the support of Shannon Treloar and Francis Hines (Department of
Sustainability and Environment) in organising the field work and processing the
Hazard Fuel field data.

References
Allen, D.M. 1974. The relationship between variable selection and data
augmentation and a method of prediction. Technometrics 16: 125–127.
Chavez, P.S., Jr., Sides, S.C., and Anderson, J.A. 1991. Comparison of three
different methods to merge multiresolution and multispectral: Landsat TM and
SPOT Panchromatic. Photogrammetric Engineering and Remote Sensing, 57:
295-303.
DNRE 1998. Forest management plan for the Central Highlands. Department of
Natural Resources and Environment Victoria.
Haywood, A. Stone, C. and Sutton, M. 2009. Estimating forest characteristics in
young Victorian ash regrowth forests using field plots and airborne laser
scanning data. Proceedings of the Biennial Conference of the Institute of
Foresters of Australia, Caloundra.
Jensen, J.R., 2005. Introductory digital image processing: A remote sensing
perspective, (3rd ed.) New Jersey: Prentice-Hall Inc.
Jeremiah, L. and Roob, R. 1992. Statement of resources, uses and values for
the Central FMA. Department of Conservation and Environment, Victoria.
Lee, A., Lucas, R. and Brack, C. 2004. Quantifying Vertical Forest Stand
Structure Using Small Footprint Lidar to Assess Potential Stand Dynamics, In:
Proceedings of the ISPRS working group VIII/2, Laser- Scanners for Forest and
Landscape Assessment, Freiburg, Germany 03-06 October 2004, pp 213-217.
Llinas, J., 2002. Centre for multisource information fusion university at Buffalo,
Bufalo, NY. Online at:
http://www.infofusion.buffalo.edu/tm/Dr.LLinas;stuff/DataFusionOverview.ppt#1
McCarthy, A.G., and Tolhurst, K.G. 1998, Effectiveness of Fire-fighting First
Attack Operations, NRE Victoria 1991/92-1994/95. Research Report No. 44
Fire Management, Dept of Natural Resources and Environment, Vic.
McCarthy, A.G., Tolhurst, K.G., and Chatto, K. 1999, Overall Fuel Hazard
Guide. Research Report No. 47 Fire Management, Dept of Natural Resources
and Environment, Vic.
Miller, A.J. 1984. Selection of subsets of regression variables. Journal of Royal
Statistical Society, Series A 147: 389–425.
Picard, R.R. and Berk, K.M. 1990. Data splitting. The American Statistician 44:
140–147.
Picard, R.R. and Cook, R.D. 1984. Cross-validation of regression models.
Journal of the American Statistical Association 79: 575–583.

18
Roecker, E.B. 1991. Prediction error and its estimation for subset selected
models. Technometrics 33: 459–468
Roff, A.M., Taylor, G.R., Turner, R., Day, M. and Mitchell, A. 2006.
Hyperspectral and Lidar remote sensing of forest fuel loads in Jilliby State
Conservation Area. Proceedings of the 13th Australasian Remote Sensing and
Photogrammetry Conference, held in Canberra, 20-24th November.
Rombouts, J., Ferguson I.S. and J.W. Leech. 2008. Variability of LiDAR volume
prediction models for productivity assessment of radiate pine plantations in
South Australia. SilviLaser 2008: Heriot-Watt University, Edinburgh, UK 8th
international conference on LiDAR applications in forest assessment and
inventory.
Rombouts, J. 2009. Experiences of the First Operational LiDAR Site Quality
Assessment in South Australia. IUFRO Division 4.01 Conference Meeting
Multiple Demands for Forest Information: New Technologies in Forest Data
Gathering. 17 – 20 August 2009
Snee, R.D. 1977. Validation of regression models: methods and examples.
Technometrics 19: 415–428.
Squire, R.O., Cambell, R.G., Wareing, K.J., and Featherston, G.R. 1991. The
mountain ash forests of Victoria: ecology, silviculture and management of wood
production. In: McKinnell, F.H., Hopkins, E.R. and Fox, J.E.D. (Eds.) Forest
Management in Australia. Surrey Beatty and Sons, Chipping Norton, pp. 38-57.
Stone, M. 1974. Cross-validation choice and assessment of statistical
predictions. Journal of Royal Statistical Society B36: 111–147
Sutherland, D., Kenny, B., Moore, P., and Ellis, P. 2006. Assessing fuel load by
remote sensing. Bushfire 2006 Conference, 6-9 June Brisbance.
Tickle, P.K., Lee, A., Witte, C., Lucas, R.M., Jones, K., Austin, J., Denham, R.
and Good, N.M. 2001. The Operational Use of Airborne Scanning Lidar and
Large Scale Photography within a Strategic Forest Inventory and Monitoring
Framework, Geoscience and Remote Sensing Symposium, 2001, IGARSS '01,
IEEE 2001 International Vol 3: 1000-1003.
Turner, R 2006. An airborne lidar canopy segmentation approach for estimating
above-ground biomass in Coastal Eucalypt forests. PhD thesis. School of
Biological, Earth and Environmental Sciences, University of New South Wales,
Sydney, NSW, Australia. July 2006.
Turner, R. 2007. An overview of airborne lidar applications in New South Wales
state forests. ANZIF Conference – Coffs harbour June 2007.
Turner, R. and Webster, M. 2005. Estimating wood volumes in Australian River
Red Gum (Eucalyptus camaldulensis) forest by combining airborne Lidar and a
digital mapping camera: A preliminary report on a local scale operational
inventory trial, In: Proceedings of SSC 2005 Spatial Intelligence, Innovation and
Praxis: The national biennial Conference of the Spatial Sciences Institute,
September 2005. Melbourne: Spatial Sciences Institute. ISBN 0-9581366-2-9

19
Appendix 1

Figure 4: Fuel layers description

20
Figure 5 Fuel Assessment Field Data Sheet

21

You might also like