spettrali
ordo di
le aree
vembre
Sensing
17-20
/amento
Sassari
lica e
Levrotto
| Image
A REGION-BASED APPROACH TO LAND-USE CLASSIFICATION OF REMOTELY-SENSED IMAGE DATA USING
ARTIFICIAL NEURAL NETWORKS
Steffen Bock
Department of Geography
University of Kiel
24098 Kiel
Germany
Commission VII
KEY WORDS: Land Use Classification Pattern Recognition Neural Networks Urban Landsat
ABSTRACT
Conventional pixel-by-pixel techniques like the maximum-likelihood method often achieve insufficient results for the classification
of intra-urban areas or complex landscape patterns on high-resolution remote sensing imagery. This is especially due to the fact, that
pixel-wise classifiers do not take into account the possible relations or similarities that may exist between one pixel and its neigh-
bours. In this paper, a method using modified co-occurrence matrices combined with a neural network was applied for the purpose
of utilizing spatial information. Instead of counting gray level co-occurrences, boundary lengths of adjacent regions were computed.
The neural network type used in this study is ATL (Adaptive Threshold Learning). ATL is a supervised feedforward network which
differs significantly in concept from the widely used backpropagation paradigm. The presented method was tested using Landsat TM
data obtained over the city of Santos/Brazil. It is shown that this approach produces promising land-use classification results in terms
of classification accuracy. In particular, the obtained land-use classes are more realistic and noiseless compared with a conventional
Bayesian method.
1. INTRODUCTION
The overall objective of image classification procedures is to
automatically categorize all pixels in an image into land-cover
classes (Lillesand and Kiefer, 1994). This is relatively easy with
conventional pixel-wise classifiers because land cover is di-
rectly related to the pixel values on an image (Gong and How-
arth, 1992b).
For the classification of intra-urban areas, however, conven-
tional pixel-by-pixel techniques like the maximum-likelihood
method often achieve insufficient results. This is due to two
facts. First, distinct urban areas represent different types of land
use. In contrast to land cover, land use is a cultural concept.
Whereas land cover is defined as the physical evidence on the
surface of the earth, the term land use relates to man's activities
or economic functions associated with a specific piece of land
(Lillesand and Kiefer, 1994). What we see on remote sensing
imagery is only the physical evidence of land use as represented
by combinations of various land-cover types (Driscoll, 1985).
Second, conventional classifiers employ only spectral informa-
tion on a pixel-by-pixel basis (Gong and Howarth, 1992a). This
strategy does not take into account the possible relations or
similarities that may exist between one pixel and its neighbours
(Gonzales and Lopez, 1992). A large amount of spatial infor-
mation is thus ignored. Therefore, accurate land-use maps
cannot be obtained through a direct transformation from re-
motely-sensed data to land-use categories; they require infor-
mation from both spectral and spatial contexts to characterize
the land use (Gong and Howarth, 1992b).
There are several types of classification which make use of
additional information, as well as the multispectral information
from a classification unit (see, for example, Mohn et al., 1987,
71
and Kartikeyan et al., 1994). In this paper, a method using
modified co-occurrence matrices in combination with a feed-
forward neural network called ATL was applied for the purpose
of utilizing spatial information.
2. METHODOLOGY
2.1 A Region-Based Co-Occurrence Matrix
One of the most popular methods to measure spatial dependen-
cies involves the use of the gray-level co-occurrence matrix.
This matrix contains the relative frequencies P;j with which two
neighbouring pixels separated by distance d and angle a occurs
on the (sub-)image, one with gray level / and the other with
gray level j. In this study, the gray-level values were replaced
by land-cover classes derived from a pixel-specific unsuper-
vised classification of multispectral imagery using the ISO-
DATA method (Hall and Ball, 1965). Furthermore, the total
length of the boundary between region i’ and j’ on the image or
within a subimage defined from a buffer zone around a region
substitutes the relative frequencies Pjj. For example, the modi-
fied co-occurrence matrix of the pattern in Figure 1 is
0 4 26
4 0 8
26 8 0
Several statistical measures, such as homogeneity, contrast, and
entropy can be computed from a co-occurrence matrix to de-
scribe specific textural characteristics of an image (Haralick et
al., 1973). But these scalar parameters contain only a part of
texture information. It is more advisable to employ the co-
occurrence matrix itself. However, it is difficult to treat two-
dimensional arrays in conventional statistical classifiers like the
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996