International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
154 
AN AUTOMATED APPROACH FOR TRAINING DATA SELECTION WITHIN AN INTEGRATED GIS AND REMOTE 
SENSING ENVIRONMENT FOR MONITORING TEMPORAL CHANGES 
Ulrich Rhein 
ISPA, University of Vechta, P.O. Box 1553, D-49364 Vechta, Germany, urhein@ispa.uni-vechta.de 
KEYWORDS: Remote Sensing, GIS, Knowledge-Base, Change Detection, Classification, Training Data. 
ABSTRACT 
Environmental monitoring analysis within an integrated GIS and remote sensing environment will be a main requirement for past and 
future projects. There is a continuously increasing synergy between GIS and remote sensing because there is a growing demand for 
spatial information, especially of high accuracy and currency. Major developments in the analysis process will be the automated 
extraction of accurate information from high resolution remote sensing imagery. New techniques, such as the use of knowledge- 
based systems, hierarchical processing, artificial neural network analysis or combinations of these techniques have shown great 
potential in a number of environmental monitoring studies. This paper presents the development of a robust classification technique, 
which is based to a large degree on standard or operational components. Any supervised classification approach needs excellent 
ground truth data to train the algorithm. It is a well-known fact that training set characteristics have a significant - if not the most 
significant - effect on the performance of any image classification. Ground truth campaigns over time series, however, can be very 
time consuming and expensive. The ultimate goal is the automated detection of changes in our environment. Regarding this goal, our 1 
strategy is not to detect areas where changes have occurred but rather those without changes. These areas will be stored in a GIS t 
database and will provide the basis for a hierarchical knowledge-based classification. This project utilises common robust e 
classification algorithms to discriminate between the major landcover types. An automated GIS process algorithm (buffer, mask, c 
overlay, etc.) will be used to eliminate or modify prospective training areas with multimodal histograms. The evaluation of the results '<■ 
in a qualitative and quantitative manner will be part of future work. c 
t 
1. INTRODUCTION 
We are modifying our environment at unprecedented rates and 
scales. We can, however, debate the specific spatial dimensions, 
rates and significance of these changes. Throughout history, 
technology has always been a key factor facilitating change. 
Current technology can create environmental changes at 
previously unknown spatial and temporal scales. Yet, it also 
offers us the ability to facilitate our investigations leading to a 
more complete understanding of human impact on our 
environment. Through appropriate use of technologies, we can 
move a significant step towards an environmentally sound 
management of the Earth’s natural resources (Ehlers, 1996). 
Classification strategies have often been highlighted themes of 
treatises. Argialas and Harlow (1990) gave an extensive 
overview of various approaches for image classification and 
image understanding within remote sensing. Beyond this, 
automation of image classification tasks is a dream of the image 
interpretation community. Various good but specific results 
were provided for example by Huang and Jensen (1997). They 
presented a machine-learning approach for acquiring 
knowledge. Another goal for various tasks, ranging from image 
acquisition to image classification, is their faster execution. The 
ultimate goal of this project is an automated change detection. 
2. DATA 
The evaluation of robust monitoring analysis methods needs 
comparably exact databases with information on landuse classes 
and their respective changes. On the one hand, our approach has c 
to deal with change detection for agricultural areas. On the I 
other hand, we deal with changes of natural/agricultural areas s 
near urban settlements. The latter means that the investigated i 
areas may be rural/agricultural with embedded settlements. 
Besides this important fact, accurate training sites and data 
should be guaranteed. The main data source for our analysis are 
Landsat-TM scenes. Other satellite remote sensing data are 
SPOT-Pan and SPOT-XS images. For special test sites, scanned 
aerial colour infrared (CIR) images are available. These aerial 
images are used to support the definition of test areas and to 
check the classification accuracy. Topographic maps at a scale 
of 1:5000 are used as the geometric basis for mapping ground 
truth areas. 
In addition, digital data from the German Authoritative 
Topographic Cartographic Information System (ATKIS) are 
used not only for the visualisation of geocoding errors but also 
for defining ground control points for the rectification process. 
The use of ATKIS as a masking layer for individual classes is 
possible in most cases. In this approach, however, this 
procedure is not applicable for all cases. For example, farmland 
and pasture areas must have a minimum size of one hectare to 
be included into the ATKIS database. It is a well-known fact, 
that any supervised classification needs excellent ground truth 
data to begin with. It has been argued, that training set 
characteristics have the most significant effect on the 
performance of any image classification (Swain and Davis, 
1978). Training data should be collected as close as possible to 
the date of remote sensing data acquisition. Table 1 gives an 
overview of our current database.