e and 
trees. 
n this 
Also 
id the 
)liers. 
ype. 
3 
> 
f high 
ilding 
image 
nplate 
model 
d tree 
first 
lution 
mated 
ocular 
and a 
high- 
lented 
use a 
iomial 
arding 
cently 
image 
ormed 
2001) 
10wed 
cision 
and 2 
metric 
ptable 
a few 
)S Pro 
1km), 
ntains 
itellite 
metric 
with 
curacy 
w data 
) ALS 
:h was 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004 
used as a 3D range data source. This has an average point 
density of 1 lidar footprint per 3-4 square metres point density. 
It was quickly established that the point density of this 3D range 
data was insufficient for reliable object model construction. 
Hence an image fusion algorithm was developed to address 
such problems. 
2. ALGORITHMS 
The overall processing steps are shown in Figure 1. The overall 
procedure consists of 3 stages. Firstly, regions of interest (ROIs) 
are defined in a focusing stage using a normalised DEM, n- 
DEM (i.e. heights above the “bare earth” terrain) and NDVI 
(Normalised Difference Vegetation Index). These “ROIs” are 
then refined using multi spectral information from the co- 
registered optical images.. Then polygons for buildings and 
ellipses for tree crown are fitted to the refined ROI boundaries 
to identify buildings and trees in the “identification” stage. 
Freres nedrrgy wa tuagger 
  
data 
= AE 
Region of = tosses i E à 
Interest for Pers f 
em S Interest for 
building emt 
structure S 
- - — 
1 
Construction of DTM from 3D range | 1 
t 
t 
1 
i 
t 
i 
1 
i 
1 
1 
i 
i 
! 
tra Ree dee i 
i 
i 
1 
2 
1 
i 
t 
i 
1 
i 
i 
i 
' 
i 
t 
i 
  
  
  
  
   
  
    
    
  
Refinement of ROI boundaries by 
multispectral and range data clues 
  
€ ZZ AU SA etae entm t ett enm 
FO LIE cumm umm cu pe ee ee ee ee — U 
  
Generalization and model fitting into 
individual objoct's boundaries 
Ferté iseauet corpus ct [ 
s saentircatton stage 
   
   
   
   
  
Building 
polygon 
Fittec! ellipse 
on tres crown 
  
  
F igure I. Overall Procedure | 
2.1 Focusing 
A focusing strategy using the n-DEM and multi-spectral 
information is commonly used. In our case, NDVI for satellite 
information and normalised colour index from ratios such as (R- 
G/R-*G) originally used for aerial photos, was adopted. A key 
point in this strategy is the use of 3D range data. There are two 
common approaches: — a so-called “Top-down” approach, 
which directly segments the 3D range data and a “Bottom-up” 
approach (Kraus and Pfeifer 1997, Axelsson 1998, Lohnnmann 
et al. 2000, Vosselman 2000), which attempts to construct a 
Bare Earth DEM (DTM). The “Bottom-up” approach usually 
produces a coarser boundary but is more suited to wide area 
applications due to a lower computational demand. 
Our “Bottom-up” strategy used a hierarchical scheme to reduce 
CPU time and update reliability in the reconstructed DTM for 
dense altitude clusters. 
The definition of a seed area is the starting point for Bald Earth 
construction. 3D range data points are re-binned using (1) to 
avoid artefacts and the local min-max detection algorithm from 
Chaudhuri and Shankar (1989) is applied to the newly binned 
height plane. 
n e ; 
h,(n,,n,) 7 max(> > h(x, v)/h,) (1) 
x=l y=l 
where h = height points at x,y coordinates 
h,=vertical renormalized factor 
h,= binned value 
n,7x/n, n,=y/n 
n = size of bin, usually at the maximum ALS data 
resolution 
From the detected local min-max points, region growing using 
the local slope (usually 25°) is used. If the dimensions of the 
827 
  
region-growing sheet is larger than the estimated maximum- 
building size, it’s likely to be a seed area (ground plane). Then 
two normal distribution are fitted to the height points on the 
"ground plane" using a window size by Kittler and Illingworth 
(1986)'s criterion. 
J(t) 2 1 2.0(P (07)log o,(t) + P(t) log o,(1)) (2) 
- 2.0(B(1)log P,.(1) - P.(r)log p.(t)) 
where P(t) : the sum of Probability Density Functions (PDFs) 
o : the standard deviation 
Now, 
If p(t) uuy(t) < the estimated object height, then p(t) is the 
mean value height of a larger object surface or it is assumed to 
be a flat ground plane so that the window size then needs to be 
extended and the estimation repeated. 
If po(t)- p(t) > the estimated object height, then p(t) is 
selected as a seed point within the window. 
Using a value of p(t) and window size, w, a gridding scheme 
can be applied. The Smith & Wessel (1990) method to 
interpolate bald earth seed points is employed here. It’s one 
kind of optimisation method for the solution of the following 
function 
(1-T) VZ-TVZ=0 (3) 
where Z : height points, T : Tension factor between 0.0 to 1.0 
When applying this method, a higher tension factor will produce 
a higher curvature surface. To construct a smoother surface at a 
lower hierarchical stage, a lower tension factor is required. In 
the first stage, the lowest tension factor and the largest window 
size are used to save CPU time. 
Figure 2 shows one example of the refinement step in the flat 
earth surface. As seen in Figure 2, the DTM detail is clearer in 
the later gridding steps and well preserves flatness. 
Now by thresholding the n-DEM (DSM-DTM), the “above 
surface points which are likely to be either trees and/or 
buildings" and “surface” points are split. Tree and building 
areas are simply separated around an NDVI value 0.3. 
Therefore, “tree ROI" and “building ROI” are defined in this 
step. 
(a) original (b) initial (d)intermediat (d) final stage 
DEM stage DTM e stage DTM DTM 
Figure 2. Hierarchical refinement of DTM 
2.2 Refinement of region of interest 
By labelling isolated areas, a set of ROIs can be defined. 
However, their boundaries are not sharp because of the poor 
resolution of the 3D range data (1pixel /3-4 m) compared with 
optical images. 
A strategy to cope with such a situation is the compensation of 
ROIs through optical image clues using mainly multi-spectral 
information. 
The overall procedure is shown in Figure 3. 
The positions of buildings and trees can be identified by 
locating *above ground" 3D range points. Thus, supervised