142
E.G.H. Gorte
space into partitions that correspond to classes. This task can be achieved by pattern
recognition (Ripley, 1996).
After the introduction, this section consists of six parts. First (a), image classification is introduced, followed by (b) an elaboration on maximum likelihood classification. Next, (c) some weaknesses are signalled and (d).impovement by refined
probability estimates is suggested. Spatial image characteristics are considered in
(e) image segmentation. Finally, (f) a case study is presented (Gorte, 1999).
7.4.1 Image Classification Methods
Reflections measured by satellite sensors depend on the local characteristics of the
earth's surface. In order to extract information from the image data, we must find
out this relationship.
Multi-spectral image classification can be an important step in the extraction of
thematic information from satellite images. The aim is to automate this as much as
possible by using suitable image processing and image analysis software. Due to the
complexity of satelite images, automatic classification can at best be regarded complementary to visual image interpretation. Section 7.4.3 will discuss the advantages
and disadvantages of computerized vs. visual methods.
In theory it is possible to base a classification on a single spectral band of a remote
sensing image (for example, on SPOT pan-chromatic), but much better results can
be obtained by using more bands (for example the three bands of a SPOT multispectral image) at the same time. Therefore, the focus will be on multi-spectral
classification.
Looking at a certain image pixel in M bands simultaneously, M values are observed at the same time. In the example of multi-spectral SPOT, where M = 3 each
pixel has three reflection values, in the green, red and infrared parts of the electromagnetic spectrum, respectively. Such a group of M numbers is sometimes referred
to as a pattern in statistics. Therefore, classification is an application of statistical
pattern recognition.
Feature Vectors and the Feature Space. In classification jargon the image bands
mentioned above are called features, because often transformations are applied to
the spectral values of an image, prior to classification. They are called feature transformations, their results derived features. Examples are: principal components, HSI
transformations, measures of texture etc.
In one pixel, the values in the M features can be regarded as components of a
M -dimensional vector, the feature vector. Such a vector can be plotted in an Mdimensional space, calledfeature space.
Pixels with similar spectral characteristics, which are likely to belong to the same
land cover class, are near to each other in the feature space, regardless how far they
are from each other in the terrain and in the image. Pixels belonging to a certain
class will hopefully form a so-called cluster in the feature space. Moreover, it is
hoped that other pixels, belonging to other classes, fall outside this cluster, but into
other clusters, belonging to those other classes.
E.G.H. Gorte
space into partitions that correspond to classes. This task can be achieved by pattern
recognition (Ripley, 1996).
After the introduction, this section consists of six parts. First (a), image classification is introduced, followed by (b) an elaboration on maximum likelihood classification. Next, (c) some weaknesses are signalled and (d).impovement by refined
probability estimates is suggested. Spatial image characteristics are considered in
(e) image segmentation. Finally, (f) a case study is presented (Gorte, 1999).
7.4.1 Image Classification Methods
Reflections measured by satellite sensors depend on the local characteristics of the
earth's surface. In order to extract information from the image data, we must find
out this relationship.
Multi-spectral image classification can be an important step in the extraction of
thematic information from satellite images. The aim is to automate this as much as
possible by using suitable image processing and image analysis software. Due to the
complexity of satelite images, automatic classification can at best be regarded complementary to visual image interpretation. Section 7.4.3 will discuss the advantages
and disadvantages of computerized vs. visual methods.
In theory it is possible to base a classification on a single spectral band of a remote
sensing image (for example, on SPOT pan-chromatic), but much better results can
be obtained by using more bands (for example the three bands of a SPOT multispectral image) at the same time. Therefore, the focus will be on multi-spectral
classification.
Looking at a certain image pixel in M bands simultaneously, M values are observed at the same time. In the example of multi-spectral SPOT, where M = 3 each
pixel has three reflection values, in the green, red and infrared parts of the electromagnetic spectrum, respectively. Such a group of M numbers is sometimes referred
to as a pattern in statistics. Therefore, classification is an application of statistical
pattern recognition.
Feature Vectors and the Feature Space. In classification jargon the image bands
mentioned above are called features, because often transformations are applied to
the spectral values of an image, prior to classification. They are called feature transformations, their results derived features. Examples are: principal components, HSI
transformations, measures of texture etc.
In one pixel, the values in the M features can be regarded as components of a
M -dimensional vector, the feature vector. Such a vector can be plotted in an Mdimensional space, calledfeature space.
Pixels with similar spectral characteristics, which are likely to belong to the same
land cover class, are near to each other in the feature space, regardless how far they
are from each other in the terrain and in the image. Pixels belonging to a certain
class will hopefully form a so-called cluster in the feature space. Moreover, it is
hoped that other pixels, belonging to other classes, fall outside this cluster, but into
other clusters, belonging to those other classes.