nhul 2004 
inet in the 
19980131. 
OBJECT-BASED EVALUATION OF LIDAR AND MULTISPECTRAL DATA FOR 
AUTOMATIC CHANGE DETECTION IN GIS DATABASES 
V. Walter 
Institute for Photogrammetry, Stuttgart University, Geschwister-Scholl-Str. 24D, 70174 Stuttgart, Germany 
Volker. Walter@ifp.uni-stuttgart.de 
KEY WORDS: GIS, Change Detection, Fusion, Classification, LIDAR, Multispectral 
ABSTRACT: 
The automatic interpretation of aerial and satellite images is one of the main important research topics in photogrammetry and image 
analysis. The aim of the interpretation is the recognition or verification of objects in images. The quality of the interpretation results 
depends among other factors on the information content of the input data. The more information in the input data the easier is the 
interpretation process. In this paper an approach is described that increases the quality of the interpretation process by using existing 
GIS data as prior information on the one hand and by combining multispectral and LIDAR data on the other hand. The approach is 
used for automatic change detection and is based on the evaluation of automatically derived training data sets from existing GIS data. 
For that reason no data dependent tuning factors have to be defined and no human interaction is necessary. 
1. INTRODUCTION 
GIS databases are very dynamic and can change very rapidly 
over the time. Automatic interpretation procedures are needed 
in order to provide up-to-date databases. One of the standard 
approaches for the automatic interpretation of aerial or satellite 
images are pixel-based classification algorithms. With that kind 
of algorithms and appropriate input data it can be distinguished 
for example between vegetation pixels and non-vegetation 
pixels. Also the pixel-wise differentiation. between different 
vegetation classes can be made very reliable. But it is not 
possible to distinguish between landuse classes that can only be 
detected by the evaluation of pixel groups and not of single 
pixels. For example it is not possible to distinguish between 
residential and industrial settlement areas only by evaluating 
single pixels. In order to distinguish between such landuse 
classes, we use an object-based classification approach that 
classifies not single pixels but groups of pixels that represent 
already existing objects in a GIS database. An n-dimensional 
vector of the feature space represents the objects. With that 
approach it is very easy to combine input data from different 
sources, like multispectral and LIDAR data or other data 
sources. 
2. EXISTING WORK 
Object based image analysis approaches for the interpretation of 
aerial and satellite images are already successfully applied to 
other problems. These approaches can be subdivided into 
approaches that use object-oriented classification rules without 
any GIS input and approaches that use exisüng GIS data to 
superimpose it on an image (also known as per-field, per-parcel 
or knowledge-based classification). 
The difference to existing approaches is that in our approach the 
object interpretation is based on a maximum likelihood 
classification and all parameters are derived from existing 
training data. In own work it was shown that the result of a 
classification could be improved significantly by the combined 
use of multispectral and LIDAR data (Walter 1999). 
Furthermore it was shown that GIS-based classification can be 
successfully used to verify different object classes in the scale of 
1:25.000 (Walter 2004). 
Object-based image analysis is also used for example in (Benz 
et. al. 2004). The basic units in this approach are also image 
segments instead of single pixels. But these segments are 
derived from image scgmentation techniques and not from 
existing databases. Therefore, this approach is more designed 
for the first acquisition of GIS objects and not for update or 
quality control. Most of the existing approaches that use GIS 
data as prior information are used for the detection and 
verification of roads (e.g. Zhang 2004) or buildings (e.g. Suveg 
and Vosselmann 2002). 
3. OBJECT-BASED CLASSIFICATION APPROACH 
In (Walter 2004) it was shown that it is possible to distinguish 
between the landuse classes forest, settlement, greenland and 
water with an object-based classification by using multispectral 
data. Because GIS databases contain typically more different 
landusc classes, this approach has to be refined. In the 
following we extend this approach and add LIDAR data as an 
input channel. This gives the possibility to evaluate more object 
characteristics. The approach will be tested on the example of 
the automatic classification of residential and industrial 
settlement objects. 
3.1 Overview 
The approach (sec Figure !) consists of two classification steps. 
In a first step, a pixel-based classification is calculated. The 
result of the pixel-based classification as well as the input 
channels (the multispectral and LIDAR data) are used as an 
input for the object-based classification that classifies not single 
pixels but groups of pixels that represent already existing