A VS BI
ve hy -
USE OF FEATURES DERIVED FROM PROPORTIONS OF CLASSES IN A
PIXEL FOR THE MULTISPECTRAL CLASSIFICATION OF REMOTE SENSING IMAGES
Ana Paula D. Aguiar
Nelson D. A. Mascarenhas
Yosio E Shimabukuro
Divisao de Processamento de Imagens - DPI
Instituto Nacional de Pesquisas Espaciais - INPE
Caixa Postal 515 - CEP 12201
Sào José dos Campos, SP - Brazil
ABSTRACT:
An analysis of the use of features derived from class proportions in a pixel for the multispectral
classification of reforested areas in Landsat images is performed. Through a Linear Mixing Model, synthetic
bands derived from those proportions are obtained either through the Constrained or the Weighted Least
Squares Procedures. The method indicates that the synthetic bands offer an alternative to the well known
dimensionality reduction techniques such as Principal Components or Canonical Analysis. Furthermore, those
bands provide a useful tool for visual interpretation, since they contain information that is related to
physical concepts (proportions) more easily assimilated than the class spectral signatures.
KEY WORDS: Mixing Model, Feature Reduction, Least Squares Methods, Classification, Image Analysis, Landsat,
Thematic.
1. INTRODUCTION
In Remote Sensing satellite images, the size of the
pixel, in general, may include more than one type
of terrain cover. When these sensors observe the
earth, the measured radiance is the integration of
the radiance of all the objects that are contained
in the pixel, implying the existence of the so-
called mixture problems.
Two main approaches have been used in the technical
literature to solve the mixture problem:
a. To improve the area estimates obtained by
conventional classification methods, which are
mainly based on the spectral characteristics of the
Pixels. In general, one class corresponds to a
single type of terrain cover. Therefore, the signal
obtained by the combination of two or more class
will not be representative of any of these classes
and, as a result, an incorrect estimation of the
area of each class is obtained. Under this
approach, one may mention the works performed by
Horwitz et al.(1975) and Ardeo(1983). The main idea
that guides this approach is to substitute the
conventional classification methods by the
estimation of the proportions of classes within a
pixel in the computation of the total area of each
class in a scene,
b. To develop methods that involve synthetic images
derived from proportions of the objects that are
contained in the image. The present work is
included in this approach, as it is described in
the following paragraphs.
Detchmendy and Pace (1972) developed a linear
mixing model to explain the variations found on the
spectral class signatures. The basic hypothesis of
such model is that these variations are mainly
caused by structural target characteristics at the
sub-pixel scale, These variations can be
interpreted as a function of the proportions of the
materials (called primary components) that
constitute scene,
As an example, according to Shimabukuro(1987), in
forested areas, three main components are found:
tree canopy, soil and shadow. Adams et al.(1990)
describe the types of land use found in Amazon
Forest, in terms of four components: vegetation,
259
Soil, shadow and wood.
Shimabukuro proposes the use of an image derived
from the shadow component in each pixel of the
image, called shadow image, as an indication of the
structure variations on the forest, that is, the
estimated shadow proportions in an image indicate
variations in age, type and shape of tree crown
cover.
The objective of the present work is to analyze the
results obtained in the multispectral image
classification (in particular through the Maximum
Likelihood criterion under the gaussian hypothesis)
with the use of synthetic bands derived from the
proportions of primary components within the
pixels. Such bands are derived not only from the
shadow component, but also from other components of
the scene.
The adopted Linear Mixing Model, the problem of
estimating the primary components proportions
within the pixels and the synthetic bands
generation are described in Section 2.
There is a relationship between the number of
features used in the classification procedure and
the corresponding computational effort. For the
Maximum Likelihood classifier, this relationship is
quadratic. Therefore, it is important to assure the
use of the minimum number of features for an
efficient classification (Richards, 1986).
Some of the feature reduction mainly used in Remote
Sensing are based on the Jeffries-Matusita (J-M)
distance and the Principal Components and Canonical
Analysis transformations (Section 3). These methods
are also useful for the visualization of color
composites of multispectral images (Richards,
1986).
The Mixing Model can be regarded as a tool to
reduce the dimensionality of the feature space to
the number of mixture components. In the experiment
presented in this paper (Section 4), the results
obtained through maximum likelihood classification
with the use of bands derived from conventional
feature reduction procedures, and these synthetic
bands derived from the Mixing Model are compared.