Multi-spectral and Multi-resolution Images for Updating Topographic Data 
Adele Sindhuber 
Josef Jansa 
Institute of Photogrammetry and Remote Sensing, Vienna University of Technology, Austria 
Comission VII, Working Group 4: Computer Assisted Image Interpretation and Analysis 
KEYWORDS: texture analysis, multispectral classification, landuse update, multiresolution images 
ABSTRACT 
Diverse remotely sensed data sets from the satelitte sensors Landsat Thematic Mapper, IRS-1C-Pan and SPOT have 
been utilized for landuse classification. In high resolution panchromatic images a texture analysis is performed with the 
help of the Förstner operator that is able to distinguish between point features, significant edges and homogeneous 
areas. In this way we classify highly textured areas such as settlements or even individual buildings. The panchromatic 
images were also used for forest segmentation by thresholding. The multispectral data were subjected to a modified 
maximum likelihood classification where each pixel has been assigned not only the class with the greatest probability 
density but also, as a second choice, the class with the second greatest probability density. For the final class decision a 
rule based approach links together the result of the texture analysis, the forest segmentation, the results of the 
multispectral classification, and, not least, certain contents the data base of the so-called Digital Landscape Model 
(DLM) of the Austrian Federal Office of Metrology and Surveying. The resulting landuse-layer has landuse-classes 
with reliability categories and a 15x15m ground resolution. A quality assessment proved that 93% of the pixels are 
identical to classes of a visual reference classification. For pixel of highest reliability the percentages grow to 9646. 
KURZFASSUNG 
Fernerkundungsdaten von den Satellitensensoren Landsat Thematic Mapper, IRS-1C-Pan und SPOT wurde für eine 
Landnutzungsklassifikation verwendet. In den hochauflósenden panchromatischen Bildern wird eine Texturanalyse 
ausgeführt, wobei der Fórstner Operator zur Anwendung gelangt, mit welchem sich punktfórmige Merkmale und 
markante Kanten von homogenen Bildbereichen trennen lassen. Auf diese Weise lassen sich hochtexturierte Gebiete, 
wie etwa Siedlungen aber auch einzelstehende Hàuser, klassifizieren. Die panchromatischen Bilder werden aber auch 
für die Segmentation des Waldes verwendet. Die multispektralen Daten wiederum werden einer modifizierten 
Maximum Liklihood Klassifizierung unterworfen, bei welcher jedem Pixel nicht nur die Klasse der größten 
Wahrscheinlichkeitsdichte sondern auch jene der zweitgrößten zugeordnet wird. Die endgültige Klassifikation wird mit 
einem Regelsystem durchgeführt, welches die Ergebnisse der Texturanalyse, die Waldsegmentation, die Ergebnisse der 
multispektralen Klassifizierung und, nicht zuletzt, bestimmte Inhalte des sogenannten Digitalen Landschaftsmodelles 
(DLM) des Österreichischen Bundesamtes für Eich- und Vermessungwesen zusammenführt. Als Ergebnis erhält man 
einen Landnutzungslayer mit Zuverlässigkeitskategorien innerhalb der geforderten Genauigkeit von 15 m x 15 m. Eine 
Überprüfung mit einer visuellen Referenzinterpretation zeigte, daß inhaltlich 93% der Pixel identisch klassifiziert 
wurden, wobei der Prozentsatz sogar auf 96% steigt, wenn man nur die zuverlässigten Klassen beurteilt. 
1 INTRODUCTION 
Providers of country-wide data sets, as for instance, the extremely high geometric quality is not a primary 
traditional national mapping offices, are facing a demand- — requirement, while the up-to-dateness is the most 
ing task today. Information needed by the customers is essential request. The modern spaceborne sensors allow 
rather specific for the respective application and varies frequent coverage of vast areas at affordable prices and 
from one task to the other, it should be as most up-to-date 
as possible and, last not least, as complete as 
economically achievable. Traditional mapping with its 
long-period update cycles of several years in the optimum 
case, but usually of decades, with its less flexible or even 
invariable and often inapproporate content is not suitable 
for modern demands at all. Secondly, the basic materials 
for deriving the data sets are still aerial photographs. 
They guarantee a very high geometric quality, but 
unfortunately at rather high costs. In many cases the 
with rather good spatial resolution. Though the spectral 
resolution is not that good at the moment, the remote 
sensing community expects the first satellites in space 
that should be able to fulfil many of the desirable. 
requirements in near future. Even with the current sensors 
a high quality classification is achievable if various 
sensors that complement each other are jointly utilized in 
a sophisticated way. Investigations in serveral countries 
concentrate on quick and economic update possibilities of 
existing data bases (Ament, 1997; Plietker, 1997). 
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 273