International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999
need for a conceptual framework was clearly expressed by
scientists as well as practitioners.
It is often written that fusion takes place at three levels in data
fusion: pixel, attribute and decision (Brandstàtter and Sharov,
1998; Csathô and Schenk, 1998; Mangolini, 1994; Pohl and
van Genderen, 1998). It presents two drawbacks. The word
"pixel" is inappropriate here ; the pixel is only a support of
information and has no more semantic significance than voxel
or «-dimension cell. Measurements or observations or signal
would be more appropriate. But even if “pixel” is corrected and
though the authors understand very well data fusion, my own
experience in teaching shows that such a categorization, also
proposed by Klein (1993) or Hall, Llinas (1997) may be
misleading and should be avoided. It may falsely imply that
fusion processes do not deal simultaneously with these different
levels. Usually, fusion of measurements results into attributes,
and fusion of attributes into decisions, but it may be different.
In Earth observation domain, one may use some features
(attribute level) held in a geographical information system to
help in classifying multispectral images (measurement level)
provided by several sensors. In this particular case, some data
are measurements of energy, and others may be symbols. Inputs
of a fusion process can be any of the levels above-mentioned, in
a mixed way, and outputs can be any of these levels. Consider
the case of the ARSIS concept which increases the spatial
resolution of a multispectral image given another image of a
better resolution not necessarily acquired in the same spectral
bands (Ranchin et al, 1996, Ranchin, Wald, 1998). It intends
to simulate what would be observed by a multispectral sensor
having a better spatial resolution. Accordingly, it simulates
measurements through a fusion process and inference models.
The formalism of Houzelle, Giraudon (1994) is preferable. It
allows all semantic levels (measurements, attributes, decisions)
to be simultaneous inputs of a fusion process. Wald (1998a)
presented several examples of this formalism applied to remote
sensing.
A search for a more suitable definition was launched with the
following principles. The definition for data fusion should not
be restricted to data output from sensors (signal). It should
neither be based on the semantic levels of the information. It
should not be restricted to methods and techniques or
architectures of systems, since we aim at setting up a
conceptual framework for data fusion. Based upon the works of
Buchroithner (1998) and Wald (1998b), the following
definition was adopted in January 1998: « data fusion is a
formal framework in which are expressed means and tools for
the alliance of data originating from different sources. It aims at
obtaining information of greater quality; the exact definition of
‘greater quality’ will depend upon the application ». (in French:
la fusion de données constitue un cadre formel dans lequel
s’expriment les moyens et techniques permettant l’alliance des
données provenant de sources diverses. Elle vise à l’obtention
d’information de plus grande qualité ; la définition exacte de
« plus grande qualité » dépendra de l’application.)
This definition is clearly putting an emphasis on the framework
and on the fundamentals in remote sensing underlying data
fusion instead on the tools and means themselves, as is done
usually. The latter have obviously strong importance but they
are only means not principles. A review of methods and tools
can be found in Pohl and van Genderen (1998), US Department
of Defence (1991).
Secondly, it is also putting an emphasis on the quality. This is
certainly the aspect missing in most of the literature about data
fusion, but one of the most delicate. Here, quality has not a very
specific meaning. It is a generic word denoting that the
resulting information is more satisfactory for the « customer »
when performing the fusion process than without it. For
example, a better quality may be an increase in accuracy of a
geophysical parameter or of a classification. It may also be
related to the production of a more relevant information of
increased utility, or to the robustness in operational procedures.
Greater quality may also mean a better coverage of the area of
interest, or a better use of financial or human resources allotted
to a project.
In this definition, spectral channels of a same sensor are to be
considered as different sources, as well as images taken at
different instants. Hence, any processing of images acquired by
the same sensor is relevant to the data fusion domain, such as
classification of multispectral imagery, or computation of the
NDVI (normalized difference vegetation index), or atmospheric
correction of spectral bands using other bands of the same
sensor. Any processing of time-series of data acquired by the
same sensor or different sensors, is a fusion process.
3. OTHER DEFINITIONS
It then has been suggested to use the terms merging and
combination in a much broader sense than fusion, with
combination being even broader than merging. These two terms
define any process that implies a mathematical operation
performed on at least two sets of information. These definitions
are intentionally loose and offer space for various
interpretations. Merging or combination are not defined with an
opposition to fusion. They are simply more general, also
because we often need such terms to describe processes and
methods in a general way, without entering details. Integration
may play a similar role though it implicitly refers more to
concatenation (i.e. increasing the state vector) than to the
extraction of relevant information.
Another domain pertains to data fusion: data assimilation or
optimal control. Data assimilation deals with the inclusion of
measured data into numerical models for forecasting or analysis
of the behavior of a system. A well-known example of a
mathematical technique used in data assimilation is the Kalman
filtering. Data assimilation is used daily for weather forecasting.
Terms like measurements, attributes, rules or decisions, are
often used in data fusion. These terms as well as others related
