International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
1.3 Research introduction
The study site is in Tokyo, Japan. Because of the high density of
population of Japan, the residential areas are full of houses. The
houses connect each other and remain almost no space between
them. Therefore, the biggest difficulty is separate the houses.
Considering the house roof generally has some ridge or peak
DSM is used to detect these local highest points which will be
used to segment the houses. We have visual light air-born
imagery with blue, green, red and near infrared spectrum. DSM
for the same area is available and rectified with the imagery.
A method of extracting and modelling 2D building outline is
proposed in this paper, which is depicted generally in part 2.
The segmentation of image-DSM pair to obtain the building
objects will be described in part 3, while the outline modelling
will be introduced in part 4. The result of the data of Tokyo
using our method will be displayed in part 5.
2. TECHNIQUE GLANCE
For super cities like Tokyo, millions of houses require accurate
yet efficient building extraction and modelling system. As we
know, DSM provides orthogonal height values of the terrain
which is not available by RGB image. It is free from color or
material reflection. It makes things simple so that makes it free
from misunderstood by non-uniform lighting, shade or shadow.
Sometimes the simple one is the effective one. The question is,
the spatial accuracy of the DSM is much lower than RGB
images. Even for radar or LiDAR derived DSM, the accuracy is
not as good as imagery because of lack of some details. For
sophisticated building modelling, DSM singular is not enough.
Efficient employing of image is necessary. Thus there are so
many literatures about the integrated using of DSM and imagery
(Rootensteiner 2003, Guo 2003, Franz 2003). We have matched
image-DSM pair at hand. We can generate the candidate area of
the buildings from DSM and project them to the image to
search more accurate features of the buildings.
We have studied a high efficient marker controlled watershed
MCW and LSNAT segmentation on DSM which can extract
both small dense houses and high or large buildings. This result
will be used here as masks of grouping building parts PPOs
derived from image. The grouping of PPOs guarantees the
accurate edge and component of a building. Based on the
building contour, a modeling process of corner extraction,
orientation estimation, and refining line model is proposed.
The technique can be described by the block diagram of Fig.1
Building contour
extraction Corner detection
Polygon modeling
bol Curvature breaks 1
[ Mask generation ] at sigma ] ( Line model ]
Curvature breaks Main orientation
[Pro segmentation | | at 2*sizma ] estimation ]
PPO grouping Midpcint refining
Region fullness intercept
processing correction
Figure 1 Block diagram of 2D building outline modelling
method
42
3. BUILDING EXTRACTION: BUILDING PPO
GROUPING
DSM only reflects height but no colour information and so
never distracts the observers by the complex texture or colour
variety for the off-terrain objects. Thus it is reasonable to use
DSM to detect buildings which are higher above their
neighbours. MCW is the best proper method for this goal.
Because the accuracy of DSM in both planar and vertical
direction is not compactable with visible image in the same
resolution, the rough building extracted from DSM only give
the approximated locations and shapes of the buildings. The
more accuracy building contour is expected to be extracted from
the image. Buildings extracted from DSM, called masks, are the
references to group the segmentation objects into buildings.
E. rs
> "boa
eCognition:;
Pure pixels. A
“Mask of
DSM
alt
Figure 2 Flow chart graphic interpretation
Small scale pixel aggregation on the image is employed to
generate a basic pure pixel object (PPO), which should only
cover one unique feature. Building blocks segmented by MCW
are taken as masks to integrate the PPO to a building footprint.
NDVI is used to remove vegetation objects from the PPO
filtered by building masks. These integrated PPOs compose a
building footprint which has quite high accuracy in terms of the
shape, as shown in Fig.2.
PPO:
Pure pixel is a kind of relative concept. Generally it refers to
pixels reflecting single attribute in a region. For example, an
area of forests which contains a certain type of tree, or area
which contains river water. In low or medium resolution remote
sensed imagery, a lot of pixels reflect mixture of the spectrum of
multiple terrain targets. In urban and high resolution imagery
case, visual interpretation can easily recognize the pure pixels.
If we define relatively small set of target, most pixels can be
taken as pure pixel belonging to a certain class, such as roof,
wall, road, road shoulder, line, bridge, tree, water, or band.
Some uncertain pixels exist, normally not for building structure.
Therefore, we can group the connecting identical pure pixels as
a uniform area called pure pixel object (PPO), and process the
image in term of these PPO to reduce the computation load and
random error caused by single pixel. To do this, we employed a
well known software of object-oriented segmentation named
eCognition in small scale threshold. eCognition merges a pair of
neighbouring objects with the least dissimilarity in sense of
spectrum and shape in each iteration, when pixel is taken as the
initial object. The whole image pixels will be grouped as many
objects of uniform features. Greater threshold gives larger
average size of the objects and less number of them. Of course,
objects under small threshold are more likely "pure". RGB
im:
thr
PP
int:
fatt
pre
CO
to :
Tre
Th
hoi
inc
ren
gre
acc
Rej
In |
obj
full
rat
un
im]
wit
sul
be
pol
pre
gra
anc
wh
4.1
Sq:
Th
ans
is
ori
col
dir
Me
the
col