2004
ype's
nt $
'"CFM
duced
ns of
> next
n and
d by
erlap
c part
of the
earer.
good
one
ng in
milar
nd as
/, the
me
imber
of the
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004 :
2.3 Identification of target shape
2.3.1 Building
The boundary defined through the SRG is relatively accurate
but still contains some undulation. There is a great deal of
published methods for the generalization of detected building
primitive from multi-spectral images and Lidar. The simplest
method is to use the Douglas-Peucker Line Approximation
Algorithm (Hershberger and Snoeyink 1992). However, a
problem using their scheme is encountered because building
shapes are irregular in our test area so that the discrimination
between noise and true building edges is very difficult. A
simple polygon approximation from Wall and Danielsson
(1984) is here applied to the refined building from the SRG and
shows satisfactory results
2.3.2 Tree crown identification
The main problem addressed here is the lack of information
required for automated individual tree detection (texture, crown
boundary) as it is missing in our 1 metre resolution imagery. In
our case, Lidar points, even though the resolution is not high
enough to precisely model or count individual trees can be
applied synergistically with the optical data. The starting point
is NDVI, because normalized colour using R-G-B scheme
doesn’t provide sufficient information for tree detection because
of the illumination effect. Also normalized colour
transformations are not useful because these are based on the R-
G-B colour space analysis. Our own colour scheme uses the
following colour spaces:
Ch 1: (R-G)/(R+G), Ch 2 : (G-NIR)/(R+NIR),Ch3 : NDVI
This algorithm is only applied to the channels within areas with
high NDVI (>0.2) values. The shadowed areas must be removed
to avoid classification problems. The definition of a shadow
mask is, fortunately, very simple. The pan-sharpened image is
transformed using the USGS Munsell HSV scheme and a k-
means classification is applied to this image. Using this method,
the shadow area can be easily defined and removed in
consecutive processing stages. The Lidar points in the high
NDVI area can be split into two parts using thresholds of n-
DEM height > 6m for points defined as trees and n-DEM height
< 0.5m for grassland points. These selected points can then be
used to provide training vectors for a maximum likelihood
classification. The classification results based on these
approaches are in extremely good agreement with trees and
grassland. Using tree masks from the classification, the Lidar
points from the tree crown areas can be re-collected. Kriging
has been established as the most reliable method to keep the
continuous change of height value in a round shape. It shows
there are relatively clear divisions between DEM peaks.
Therefore the key issue is how these DEM peaks can be divided.
Wood (1996) studied the detection of topographic features from
DEMs. Usually, a sloping surface that is concave in the cross-
sectional direction is a channel so the channel points of a tree
DEM can be detected using Wood's method. Each tree crown is
enclosed by a detected channel point so we can split the tree
crown by removing weakly connected components using
channel points. In each tree DEM patch, ellipse fitting can be
applied using Pilu et al. (1999)'s method. Then the eccentricity
is checked. If this value is higher than some threshold, in the
weakly connected parts, one more split is made and the ellipse
is re-fitted. This process is continued until patches are fitted or
have fewer points than a given lower threshold.
829
3. RESULTS & ASSESSMENTS
The final three products consist of - DTM, boundaries of
building structures and ellipses representing tree crowns. These
products have been verified using GIS data (OS®
MasterMap®) and visual inspection.
3.1 DTM
A DTM covering an area of 8 by 8 km was constructed (Figure
6) in 4 stages using the hierarchical gridding scheme for the
entire east London area. A quality assessment comparing
commercial DTMs (such as the 50m OS® Panorama®) is
impossible because of the large resolution differences but visual
inspection in detail shows large landscape objects like the
Millennium dome can be successfully removed and small
natural alternations in height are well preserved.
Thames around the
Millennium dome
(a) Constructed DTM for the
whole East London area
(c) Constructed DTM
Figure 6. Constructed DTM
3.2 Building boundaries
The Building Detection Metrics (Shufelt & McKeown, 1993)
scheme is used here to evaluate the accuracy of building
outlines compared with OS® MasterMap® data. In the Shufelt
& McKeown scheme, quality assessment factors are defined as
below.
Building Detection Percentage = 100 TP /(TP+FN)
Branching Factor = FP / TP
Quality Percentage = 100TP / (TP + FP + FN)
where TP: True Positive (Both data sets (detected building and
OS data) classify the pixel as being part of an building)
TN: True negative (Both data sets classify the pixel as being
part of the background)
FP: False Positive (Detected data set classifies the pixel as a
building, OS data set classifies it as background)
FN: False Negatives (Detected data set classifies the pixel as
background, OS data set classifies it as a building)
Figure 7 (a) and Table 1 shows the detection accuracy of
building boundaries in a 1.0 by 1.5 km area using the method
described here. As seen in Figure 7 (b), small building
structures such as houses and irregular shaped buildings are
successfully detected.
Building Detection Percentage 74.97 %
Branching Factor 0.19
Quality Percentage 65.67 %
Table 1. Quality assessment of detected building boundaries cf.
Parish et al., (2003)
