y the
GISs
lating
shing
scale
. The
rithm
final
ition,
natic.
ca, or
g 3D
iency
nent.
and
F this
a 2D
Vays:
dings
tes a
non-
icles.
ping
The
areas
ed at
, the
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
Spatial Information Systems and National GIS for control,
management and planning, necessitated the development of full
or partial automation methods for establishing and updating
these systems. In the first system, the aim is to achieve a
complete 3D visual presentation of the buildings. This is
generally accomplished by using a large scale (—1:4000) at a
high resolution. The methods of automation for this system are
usually based on model fittings (Building Model Schemes)
divided into four categories (Tseng and Wang, 2003):
polyhedral models, prismatic models, parameterized polyhedral
models and CSG models. In contrast, the second system aims to
achieve mapping of the 3D building contour (outline) as viewed
from the top. Such mapping is mostly accomplished by using a
medium scale (~1:40,000) and medium resolution. The methods
of automation for this system vary and differ at the automation
level offered. Usually, the automation level is determined by
the point of origin. In the automatic methods the initial votes or
the rough location of the buildings are automatically extracted.
In some methods, the initial votes or the initial rough locations
arc 3D when extracted by exploiting the DSM or DEM
(Weidner and Forstner, 1995; Cord and Declercq, 2001; Ruther
et al, 2002). In other methods, they are 2D when using
classification or texture analysis (Kokubu at al, 2001), shadow
analysis (Irvin and McKoewn, 1989) or finding local
maximums in a cumulative matrix of possible votes (Croitoru
and Doytsher, 2003). In the semi-automatic method the voting
is accomplished manually. Relying on the initial votes or the
initial rough location the building contours are extracted in the
image space. Michel et al. (1999) suggest that the initial vote
would be 2D (i.e., on the left image) and performed manually.
The rough location would be spotted by using Region Growing
operations on the intensity and disparity images. The exact
location and the matching of the images would be carried out
using Hough Transform or Snake, according to the shape and
the operator’s decision. Ruther et al. (2002) focus on building
mapping in informal settlement areas and suggest extracting the
rough location from the DSM. The exact location is extracted
from an orthophoto using the Snake method. We suggest a new
approach to mapping the buildings layer for GIS systems. This
approach will facilitates a semi-automatic 3D building
extraction from medium scale images within a nonstereoscopic
environment and without using 3D spectacles. This is done by
relying on an initial manual 2D vote.
1.3 Uniqueness of the Research
Usually it is accepted to divide the aerial images into small,
medium or large image scales ranging from 1:70,000 to 1:4,000
(Mayer, 1999). This research is unique because it focuses on
medium scale (—1:40,000) panchromatic aerial images. There is
a significant difference between large (—1:4,000) and medium
(~1:40,000) scale images. In the first, the buildings appear
clearly, the DSM extracted from those images is detailed, and
describes the surface in a credible manner. Therefore, the
accuracy of the resulting 3D mapping of the building space is
high. In the second, the buildings do not appear clearly, the
DSM is less detailed, and describes the surface in a unreliable
manner. In this case, the accuracy of the 3D building space
mapping is low and it is necessary to map the building contour
(outline) only as it appears in a top view.
775
The considerations for focusing on medium scale images were
as follows: at this scale, the images include a vast area which is
useful for mapping and updating a large area, and the National
GIS is constructed on this scale (Israel-1:40,000, France-
1:30,000, etc) and moreover, this scale will soon be available in
commercial satellite images. In addition, while there are many
studies dealing with large scale automatic mapping, there are
fewer studies dealing with the medium scale.
2. METHODOLOGY
2.1 The Algorithms
The inputs of a semi-automatic system for building mapping are
two aerial images with an overlapping area. The algorithm
consists of five stages (Figure 1): pre-processing, left image
operations, height extraction, right image operations and
mapping the object space. The first stage is performed manually
in order to achieve three purposes, namely, a mathematical
solution of the model, image processing and a manual vote on
the desired building roof. From the second stage onwards, the
process is fully automatic. The second stage includes extraction
of the building contour in the left image space. The height is
calculated at the third stage by finding the homologue points (in
the right image) for each of the left contour vertexes. Now the
initial vote can be transferred to the right image space. The right
image-building contour is extracted at the fourth stage in the
same way as the left contour was extracted in the second stage.
At the last stage, the final 3D building contour is calculated
using the information achieved in the previous stages.
2.2 Stage 1: Pre-Processing
In order to facilitate a semi-automatic process it is necessary to
prepare the environment. This includes scanning and saving the
images, finding the model's solution and performing operations
on the images to emphasize the mapping object in relation to its
background and to achieve radiometric proximity (calibration)
between images. At this point, the operator can vote (point on)
the building roof. Pointing on the building is performed on the
left image within a nonstereoscopic environment and without
using spectacles. The input consists of image coordinates of any
point on the building roof N, = (ap, This manual
operation defines the level of automation as semi-automatic.
From this point on, the process is fully automatic.
2.3 Stage 2: Left Image Operations
At this stage, the 2D building contour on the left image is
extracted by using the Region Growing method (Eq.1). Where:
1 is the average of radiometric values around the vote. o is the
standard deviation of radiometric values around the vote and
p is the factor of the standard deviation.
n; = (x, y.)
if AA pH (1)
then n, € roof _ building
