Symposium on Remote Sensing for Resources Development and Environmental Management / Enschede / August 1986
Comparison of classification results of original
and preprocessed satellite data
Barbara Kugler & Rüdiger Tauch
Department of Photogrammetry and Cartography, Technical University of Berlin, FR Germany
ABSTRACT: The production of satellite image maps is one of various objectives of image processing. Satellite image maps
are very important tools as orientation maps in regions, where other topographic maps are either not available or out of date.
They may also serve as basic information for thematic maps for landuse, geological and ecological purposes.
The methods of multispectral classification are applied to obtain landuse maps from satellite image data. These methods
are based on the characteristic spectral reflectance of objects that enables to extract feature sets for different classes. By
means of the feature sets it is possible to assign every picture element (pixel) to one of the various classes. For digital image
classification the spectral signatures are the substantial source of information. Therefore usually roughly preprocessed data
are used. The classification of preprocessed data offers advantages if the concerned area is covered by several satellite
image scenes. In this case the data undergo radiometrical and geometrical processing before classification takes place.
The purpose of this paper is to investigate digitally classified original and preprocessed data. The classification results of
the two data types are compared.
1 INTRODUCTION
During the last years many remote sensing techniques were
developed. The satellite-scanned-data is used as basic
information for many applications like the production of
image and thematic maps for urban planning, land resource
management such as agriculture, etc. All applications
require several stages of processing until satisfying results
can be obtained. For the production of image maps it is
necessary to take regard of radiometric and geometric
effects, which are caused by sensors or spacecraft. There
fore digital methods are developed, which take account of
these disturbances.
Radiometric corrections such as contrast enhancement,
corrections of pixel-and line-interferences, destriping etc.
improve image quality and are prerequisite for further
steps of processing.
In order to compile maps with high accuracy it is
essential to relate the satellite-scanned-data to the ground
coordinate system. For a comparison of data from
different dates or sensors the geometric relation to each
other must be guaranteed. These problems can be solved
with rectification and resampling algorithms.
For other purposes like landuse or forest damage investi
gations multispectral classification is used in order to gain
the desired information. The digital methods of image
classification are based on the spectral reflectances of
objects. In the images, these spectral reflectances are
represented through the grey-values in every spectral
channel of each pixel. Every kind of radiometric and
geometric image processing methods influence these values
in a different way, with the consequence that the feature
sets for the different classes change. This causes an
alteration of classification results. For that reason normal
ly roughly preprocessed data is used. But sometimes it is
advantageous or even necessary to deal with better prepro
cessed data. For a supervised classification a contrast
enhancement in every spectral channel makes the inter
active definition of training areas easier. The multitempo
ral or multisensoral classification assumes already rec
tified data. Otherwise it is not possible to assign picture
elements, which represent the same ground area, to each
other. In the case of the multitemporal data the different
dated images do not match together identically, in the case
of data from different sensors there are possibly differ
ences in the image resolution. All these influences on grey-
values caused by preprocessing make it necessary to inves
tigate classification results of data, which is preprocessed
in different ways.
In the following some usual image processing methods
are described briefly. After processing the digital data
with linear contrast enhancement and/or rectification and
resampling with the methods of nearest neighbourhood and
bilinear interpolation, the obtained images serve as input
material for supervised multispectral classification. The
classification results and their comparisons are presented
subsequently.
2 CONTRAST ENHANCEMENT
Normally satellite-scanned-data are available as digital
data on CCT's. The quality of the original data is often
very poor. The images are dark and have only low con
trasts, which make a visual interpretation difficult. There
fore the first step of preprocessing is a contrast enhance
ment for every channel separately. The distribution of the
grey values of original data only takes a small part of the
possible values between 0 and 255. A linear contrast
enhancement allows to stretch the values within the whole
possible interval. The results are better in contrast and as
a whole in the image quality. Other radiometric prepro
cessings are often necessary; however our investigations
are restricted to the effect of a linear contrast enhance
ment.
3 RECTIFICATION AND RESAMPLING
For many applications it is necessary to get a relation
between the satellite-scanned-data and the map grid. This
can be achieved by a rectification and resampling of the
data. The rectification can be realized with the help of
transformation polynomials, where the coefficients are
evaluated by a least squares adjustment. For the compu
tation of the grey-values of the new pixels the following
resampling algorithms are possible: nearest neighbourhood,
bilinear interpolation or cubic convolution. The last two
methods determine the new values by considering the
surrounding pixels, with the result of artificial grey-values,
whereas the first method takes only values, which already
exist in the original image.
Some investigations (Kahler, Milkus, 1986) show, that a
row- and column-doubling of the image before rectification
and resampling improves the result, especially the contrast
of small linear features. This means that the resolution has
been artificially enlarged. One of our case-studies investi
gates wether such a data handling will be meaningful for
classification of rectified images.
