COLOUR-BASED AND CRITERIA-BASED METHODS FOR IMAGE FUSION
Oguz Gungor, Jie Shan
Geomatics Engineering, School of Civil Engineering, Purdue University, 500 Stadium Mall Drive,
West Lafayette, IN 47907, USA - {ogungor, jshan}@ecn.purdue.edu
Commission VII, WG VII/6
ABSTRACT:
Panchromatic and multispectral images are useful for the acquisition of geospatial information about the Earth surface for the
assessment of land resources and environment monitoring. Panchromatic images usually have a better spatial resolution than the
multispectral ones of the same sensor, while the multispectral images provide spectral properties of the objects. Image fusion
methods are needed to find the missing spatial details in the multispectral images using the panchromatic ones and transfer these
details into the multispectral images without or with limited spectral content distortion. This study addresses two classes of image
fusion approaches: colour-based methods and statistical methods. Specifically, the traditional RGB to IHS transform is generalized
from 3-D to n-D such that it can handle multiple image bands. As for the statistical methods, we propose a criteria-based approach
that produces fusion products to meet a set of predefined desired properties. Principles, solutions, and formulation regarding these
two approaches are presented. The proposed methods are tested with QuickBird images. Fusion results are evaluated visually and
quantitatively with discussions on their properties.
1. INTRODUCTION
Remote sensing sensors capture the energy reflected or emitted
from the objects and convert it into the digital numbers to form
images. A sensor has a fixed signal to noise ratio associated to
hardware design. An object can be detected only if sufficient
amount of energy reaches the sensor. The energy to be collected
by the sensor is related, among others, to IFOV (instantaneous
field of view) of the sensor and the capability of the sensor to
collect the energy over a certain spectral bandwidth.
IFOV of the sensor is inversely proportional to the spatial
resolution of the image collected. The larger the IFOV, the
lower the spatial resolution, since a sensor with a larger IFOV
collects energy from a larger area on the ground. On the other
hand, the amount of energy that reaches the sensor can also be
increased by collecting the energy over a broader spectral
bandwidth. This means that reducing the IFOV and increasing
the capability of the sensor to collect energy over a larger
spectral bandwidth may retain the spatial resolution of the
image (Pradhan,2005). Panchromatic sensors collect the energy
reflected by the objects over a broader spectral bandwidth with
a narrower IFOV; therefore, panchromatic images have more
spatial detail content than the multispectral images of the same
sensor. This is why panchromatic images usually have a better
spatial resolution than the multispectral images of the same
sensor.
Image fusion intends to enhance the spatial details in the
multispectral images by using the panchromatic ones. Over the
last two decades various image fusion algorithms have been
introduced (Pohl and van Genderen,1998). These methods are
designed to accomplish two main tasks: extract the spatial
details from the panchromatic image, and transfer them into the
multispectral image using certain fusion rule or transform.
Among the various types of image fusion methods, the study
will address the colour-based methods and statistical methods.
The most representative colour-based approach is based on the
RGB to IHS transform. The RGB colour space is ideal for
colour image generation. Images are displayed on monitors
using RGB colour system and most image processing
algorithms use RGB colour space for image processing
applications. However, it has limitations (Gonzales and
Woods,2003). The RGB colour space is not intuitive and not
practical in colour selection. It is almost impossible to
distinguish a colour from another only with RGB colour
coordinates. In addition, it is device dependent. Different
monitors and even an adjustment to the same monitor give
different results. On the other hand, IHS colour space has a
significant advantage over RGB colour space. IHS colour space
makes possible to manipulate each colour attribute individually.
In a multispectral image, spatial content is retained in the
intensity component (Chibani and Houacine,2002) and spectral
information is preserved in its hue and saturation components
(Pohl and van Genderen,1998; Gonzalez-Audicana et al.,2006).
Using IHS colour space, the intensity component of an image
can be changed without modifying its hue and saturation
components. Due to this property of IHS colour space, it has
been an ideal tool for image processing applications such as
contrast enhancement and image fusion where the goal is
enhancing the spatial content of the image while preserving its
spectral properties.
Statistical image fusion methods transfer the spatial detail from
the panchromatic image using the statistical properties of the
input panchromatic and the multispectral images. These
methods include principal component analysis (PCA), linear
regression method (Price, 1999), spatially adaptive image fusion
(Park and Kang, 2004) and a-p methods (Gungor and
Shan,2005, 2006), all of which have clear statistical
interpretations.
In this study, we will present two approaches, respectively, in
the categories of colour-based fusion methods, and statistical
fusion methods. In the colour based approach, the classical
RGB-IHS transform will be generalized from 3-D space to n-D
space so that it can handle any number of bands of multispectral
images. It is shown that the generalized IHS transform is
1057