1089
FUSION OF REMOTELY SENSED MULTISPECTRAL IMAGERY AND LIDAR DATA
FOR FOREST STRUCTURE ASSESSMENT AT THE TREE LEVEL
S. S. Ali 3 ’ 1 *„ P. Dare b , S. D. Jones c
d b School of Geography, Population and Environmental Management, Flinders University, Adelaide, South Australia
- (sohel.syed, paul.dare)@flinders.edu.au
c Mathematical & Geospatial Sciences, RMIT University, Melbourne - simon.jones@rmit.edu.au
Commission VII, WG VII/6
KEY WORDS: Airborne remote sensing, Classification, Digital photogrammetry, Fusion, Forestry, Lidar, Multispectral imagery
ABSTRACT:
A new feature-level fusion is presented for modelling individual trees by applying watershed segmentation and subsequent
classification, using tree heights and tree crown signatures derived from light detection and ranging (lidar) data and multispectral
imagery. The study area is part of the Moira State Forest, New South Wales, Australia where the dominant tree species are native
eucalypts. In this study, airborne lidar data and four band multispectral imagery were acquired. A digital surface model (DSM) was
generated from the lidar first return data and a digital terrain model (DTM) was derived from the lidar last return data. A tree crown
model was computed as the difference between DSM and DTM using appropriate height thresholds. A marker-controlled watershed
segmentation algorithm was used to extract individual tree crowns from the lidar data. The resulting crown polygons were overlaid
on the four band multispectral imagery to extract the spectral signatures of the tree crowns. A principal components transformation
was applied to the four-band dataset to replace the highly correlated original bands with those of reduced correlation. In addition,
two lidar derived texture and height layers were included in the fusion procedure. The application of the maximum likelihood
technique led to a high classification accuracy. An average classification accuracy of 86 percent was achieved and this procedure
outperformed the original four-band maximum likelihood classification by 23 percent. The success of the tree crown extraction
algorithm in old growth areas was higher than in more juvenile areas where the crowns were more scattered. It was also observed
that large crowns were better delineated than small ones. The results indicate that this fusion modelling strategy may prove suitable
for estimating and mapping the crown area, height and species of each tree.
1. INTRODUCTION
1.1 Motivation
Individual tree components (both in the horizontal and vertical
plane) are important parameters for developing a better
understanding of how forest ecosystems function. Lidar data
provide accurate measurements of forest structure in the vertical
plane; however, current lidar sensors have limited coverage in
the horizontal plane. Conversely, high resolution multispectral
imagery provides extensive coverage of forest structure in the
horizontal plane, but is relatively insensitive to variation in the
vertical plane. Therefore, it is desirable to synergistically use
both sensors for mapping forest parameters at a fine scale.
Delineating individual trees and extracting relevant tree
structure information from fused remotely sensed data have
significant implications in a variety of applications such as
reducing fieldwork required for forest inventory (Gong et
a/., 1999), assessing forest damage (Kelly et al.,2004) and
monitoring forest regeneration (Clark et al. ,2004).
1.2 Imagery and lidar data fusion
There have been several attempts to fuse lidar and high spatial
resolution imagery for individual tree attributes collection
(Baltsavias,1999; Leckie et al.,2003). The strong argument of
fusion is that the lidar measurements do not distribute
homogeneously and usually have gaps between them. As a
result, the three-dimensional structure of the objects might not
be very well defined (Baltsavias,1999). It thus becomes fairly
complex to obtain a good 3D model of the canopy architecture
of each tree with a low density of lidar returns. The idea of
exploiting the complementary properties of lidar and aerial
imagery is to extract semantically meaningful information from
the aggregated data for more complete surface description.
Sua'reza et al.(2005) propose a data fusion analysis with lidar
and aerial photography to estimate individual tree heights in
forest stands. The tree canopy model is derived from lidar
layers as the difference between the first pulse and last pulse
return. Information about individual trees was obtained by
object-oriented image segmentation and classification. This
analysis provided a good method of estimating tree canopies
and heights. However, the method of segmentation and
classification are too image dependent. The classification
parameters were not defined automatically and exhibit no clear
relationship to allometry factor. Instead, they were defined
empirically following a trial-error process. Leckie et al. (2003)
applied the valley following approach in to both lidar and
multispectral imagery and found that the lidar can easily
eliminate most of the commission errors that occur in the open
stands while the optical imagery performs better for isolating
trees in Douglas-fir plots.
This study attempts to use a new feature-level fusion
methodology for modelling individual trees. The method
* Corresponding author.