Page 292 - Digital Analysis of Remotely Sensed Imagery

254 Cha pte r Se v e n

analyst has access to or is about to analyze. They form most of the
input fed into the computer. Virtually, these remotely sensed data
represent values or DNs of pixels, usually in the multispectral domain.
They capture the ability of ground objects to reflect or emit energy at
various wavelengths.
Information, on the other hand, refers to the final outcome derived
from the analysis of the data. As a type of specially processed data,
information is able to provide answers to questions related to these
data. In other words, information is the data useful for a particular
application. A very important difference between the input satellite
data and the classification results that are regarded as the information
derived from the data is the range of pixel values and their meaning.
Raw satellite data may have a pixel value ranging from 0 to 255, the
exact values being determined by the quantization level of the sens-
ing system. All pixel values are indicative of the amount of reflective/
emissive radiation. In contrast, the pixel value range of a classified
image is much narrower, usually numbered no more than tens. The
objective of image classification is to convert such a vast amount of
data into useful information. During the conversion, a wide range of
DNs is reduced to a certain number of codes, each ideally correspond-
ing to a meaningful ground cover in the classified results. Therefore,
image classification is essentially a process of data generalization
during which a range of pixel values corresponding to a ground cover
is amalgamated into a single code. How these pixel values should be
amalgamated depends on the classifier used and how many informa-
tion codes are preserved in the classification outcome. These codes are
collectively known as information classes. They are usually rendered
graphically in map format with the desired geometry (e.g., georefer-
encing maps) toward the end of image analysis.

7.1.4 Spectral Class versus Information Class
A spectral class is defined as a cluster of pixels that are characterized
by a common similarity in their DNs in the multispectral space.
Whether a group of pixels can be regarded as one cluster is subjective,
dependent on the specification of spectral distance among these pix-
els. If the distance between a pixel and a group of pixels falls within
the specified threshold, this pixel is considered a part of that cluster.
An information class, on the other hand, is a category of ground fea-
tures retained in the classification results. Every category included in
the enacted classification scheme represents an information class. It
corresponds to a specific type of ground cover or feature to be
extracted from remotely sensed data. The purpose of image classifi-
cation is to map the input data into these information classes as accu-
rately and reasonably as possible.
There is a complex relationship between information classes and
spectral classes. Rarely, they correspond to one another neatly. In