ISPRS, Vol.34, Part 2W2, “Dynamic and Multi-Dimensional GIS", Bangkok, May 23-25, 2001 
103 
lation, the 
old for the 
ed by the 
lent of 3S 
cientifically 
jld actively 
)int for the 
je people's 
ntemporary 
Technical 
39 
Nide Area 
f Surveying 
tection and 
House of 
AUTOMATIC REGISTRATION OF SATELLITE IMAGE TO MAP 
Kensaku FUJII 
NTT Cyber Space Laboratories 
1-1 Hikarinooka Yokosuka-shi, Kanagawa 239-0847, Japan 
Tel:+81-468-59-2431, Fax:+81-468-59-2829, E-mail:fujii@marsh. hil.ntt.co.jp 
Tomohiko ARIKAWA 
NTT Cyber Space Laboratories 
KEY WORDS: Registration, Satellite image, Vector map, Normalized Difference Vegetation Index, Generalized Hough Transform 
ABSTRACT 
This paper introduces a fully automatic method for registering satellite images to vector maps. The key idea is to automatically match 
the man-made objects extracted from satellite images against the corresponding objects in maps. The proposed method consists of 
two procedures: feature extraction for matching, and mismatch determination. 
Matching features are extracted using indicators derived from multi-spectral data of satellite images. The normalized difference 
vegetation index is used as a spectral indicator. It is well known to provide the accurate differentiation of vegetation from man-made 
objects. 
In order to minimize registration mismatch, we apply the voting technique based on the generalized Hough transform. Voting is used to 
judge the matching levels of displacement candidates. Mismatch is determined by voting using the positions of the features and those 
of topographical objects projected onto satellite images. 
Our approach is validated in experiments using actual multi-spectral date of satellite images and standard topological maps. The 
experimental results show that the registration is extremely accurate; the registration mismatch is just a few pixels. This level of 
performance confirms that our approach can automatically register satellite images against maps. To quantify the average 
performance of the proposed approach, we also tested the approach of using image edges as matching features. These results reveal 
that our approach yields high accuracy and small mean square error. It is clear that our approach can extract effective features from 
satellite images because it makes good use of topographical reference objects through voting. 
1. INTRODUCTION 
The registration of airborne and spaceborne images plays a key 
role in the field of photogrammetry. The demand to automate 
such registration to maps is rapidly increasing in many 
applications such as agriculture, town planning, map making 
and updating, and disaster prevention. 
Various sources can be used for GIS (Geographical Information 
System) applications. A very recent source for these 
applications is to use satellite images acquired by commercial 
satellite systems (Space Imaging), (ORBIMAGE). Recent 
progress in spaceborne sensing methods has made it possible 
to be easily acquired large amounts of data that offer excellent 
resolution (approximately a few meters). Therefore, satellite 
images are now one of the most important sources in GIS. 
Many approaches have been presented for the semi-automatic 
or automatic interpretation of satellite images (Vogtle & Steinle, 
2000), (Klang, 1997). With regard to registration to maps, 
however, several problems remain. 
The problems fall into two classes: problems with geometry 
mismatch and problems with image processing, i.e. feature 
extraction from satellite images. The former is mainly due to the 
fact that satellite images have some errors in geometry caused 
by mechanical device limitations or measurement error. With 
the rapid development in remote sensing, highly accurate 
geometry can be achieved by direct geo-referencing. However, 
the best level of position alignment accuracy equals that of 
1/25,000 scale maps without Ground Control Point (GCP) 
correction. Since 1/2,500 scale maps are now being used as 
GIS sources, the use of GCP will always be needed for 
registration. 
Due to the characteristic of satellite images, it is difficult to 
process images i.e. feature extraction (Vohra & Dowman, 2000), 
(Hild et at., 2000). Digital image processing becomes an 
important tool for the quantitative and statistical analysis of 
remotely sensed images (Ebner et al., 1999), (Grun & 
Baltsavias, 1997). However, such images often contain 
complex natural scenes and their resolution is at most a few 
meters. This seriously impacts feature extraction, and 
erroneous extraction seriously degrades registration accuracy. 
It is necessary to achieve the robust interpretation of such 
images. 
This paper describes a fully automatic method for the 
registration of satellite images to vector maps. The method 
provides the automatic analysis of satellite images and objects 
in maps, and completely eliminates the need for a priori 
information. The general idea is to automatically match man 
made objects extracted from satellite images against the 
corresponding objects in maps. The proposed method consists 
of two procedures: feature extraction for matching, and 
mismatch determination. Our approach to identifying man 
made objects is based on an indicator analysis of multi-spectral 
data of satellite images. The normalized difference vegetation 
index is used as a spectral indicator. It allows vegetation areas 
to be differentiated from man-made objects (Schilling & Vogtle, 
1996). Additionally, in order to minimize registration mismatch, 
we apply the voting technique based on the generalized Hough 
transform. We also present the results of automatic registration 
experiments using actual satellite images and vector maps. 
This paper is arranged as follows. Section 2 briefly explains the 
procedure of matching feature extraction. Section 3 details the 
process of determining registration mismatch. The results of 
some experiments are presented in section 4. Finally, we 
conclude this paper with future directions of our research. 
2. MATCHING FEATURE EXTRACTION 
This section describes our approach to extract matching 
features from satellite images. The general idea is that the 
features need to be correspond to map objects, which are 
usually man-made objects, for example building, house and so 
on. In detail, we use the NDVI (Normalized Difference 
Vegetation Index), which is well known to accurately differentiate