in: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
COMPARISON OF METHODS FOR AUTOMATED BUILDING EXTRACTION FROM
HIGH RESOLUTION IMAGE DATA
G. Vozikis
GEOMET Ltd., Faneromenis 4, 15561 Holargos-Athens, GREECE
george.vozikis@geomet.gr
KEY WORDS: Photogrammetry, Building , Detection , Transformation, Model, Pattern
ABSTRACT:
This paper discusses a comparison analysis of different methods for automated building extraction from aerial and spacebome
imagery. Particularly approaches employing the Hough Transformation, Pattern Recognition Procedures and Texture Analysis are
examined. Throughout this investigation advantages and disadvantages of the mentioned methods are examined, in order to see
which procedures are suitable for extracting the geometric building properties, and thus to automatically create a DCM (Digital City
Model). The examined data sets consist of panchromatic imagery coming from both very high resolution satellites, as well as line
scanning aerial sensors. A quantitative and qualitative assessment will help to evaluate the previously mentioned procedures.
1. INTRODUCTION
Automated building extraction from high resolution image data
(either airborne or spacebome) is becoming more and more
mature. Everyday new techniques are investigated and the
results are getting more and more reliable, while the degree of
automation increases. Each building extraction method is of
course coupled to certain pros and cons. The use of the Hough
Transformation has proven to be a very promising tool in the
frame of the automated creation of Digital City Models
(DCMs), by extracting building properties from optical data.
But also approaches based on Image Matching or Texture
Analysis seem to provide usable results. A DCM is described
through the outlines of buildings outlines of an urban area.
Vertical walls are assumed, and the elevation information of
these buildings can be taken from a DSM (Digital Surface
Model). The creation of the DSM and the assignement of the
elevation value is not discussed in this paper, thus when
mentioning DCMs we actually mean the Model that holds the
2D outline-information of a building.
The goal of this paper is to conclude for which kind of data sets
and accuracy pretensions a certain approach is recommendable.
Moreover, the reachable degree of automation is also examined,
in order to see how reliable results are that were produced
without human interaction.
Table 1 : Examined data sets.
Sensor
Location
GSD (m)
Extents (km)
ADS40
Valladolid, Spain
0.25
1 x 1
HRSC-AX
Bern, Switzerland
ca. 0.3
0.2 x 0.3
Quickbird
Denver, USA
0.6
16.9 x 16.5
IKONOS
Athens, Greece
1
9.7 x 12.3
Orb view 3
Orange, USA
1
0.6 x 0.7
Altogether, five different datasets, coming from airborne and
spacebome sensors, were examined. These datasets depict
urban regions with varying building sizes, patterns and
densities. It should be mentioned here that only subsets have
been used for the investigations.
2. DESCRIPTION OF WORKFLOWS
2.1 Hough Transformation
The proposed workflow for automated building extraction from
image data by employing the Hough Transformation has been
thoroughly described in Vozikis (2004). Figure 1 shows the
major steps of the process.
Figure 1: Proposed workflow for automated building extraction
from image data
All steps in this workflow are highly automated and human
interaction is reduced to a minimum.
In the following the 4 major steps are briefly described.
