unction
success
in the
) semi-
raction,
he most
of roads
del was
ometric
inctions
, 1995).
feature
ented in
e aerial
spectral
der and
system
out the
les. The
nsity or
area for
past. It
objects,
analysis
building
'erfectly
1 (1991)
building
eed for
porating
(traction
ted such
dataset
y place
gorithm
ispectral
iques in
standing
ral data
Remote
bility of
5 on the
level of
| image
pretation
ing base
1 and
etability,
sed or
ition on
ition can
5). Reed
entation
entation
oodcock
1998).
Generally speaking, feature extraction in the remote
sensing context is the identification of the subset of the
spectral bands that are most useful for converting spectral
data into information about the scene. Therefore most
feature extraction methods rely on high spectral
resolution, aspatial, field and laboratory data (Warner, et.
al., 1996). Recognition of shapes has not been considered
in remote sensing as extensively as it has been in
computer vision and photogrammetry because the
resolution generally available in the past has been
insufficient to define shape with any degree of precision
(Richards, 1993, p130). Thematic accuracy has been a
primary measurement in remote sensing image
processing. With the advancement in remote sensing
technology, high resolution data are now becoming
available, and boundary tracking accuracy or the
geometric accuracy of classified objects will be a
significant aspect in remote sensing image processing.
Johnson and Howarth (1987) undertook a simulation
study on the effects of resolution on land cover and land
use theme extraction using airborne data. The results
indicated that as spatial resolution increases, high spatial
frequency land cover classes are extracted in increasing
detail while only the precision of border location
improves for low spatial frequency land cover classes.
The class or thematic accuracies are similar at different
spatial resolution. High resolution data are not necessary
for applications which focus on the classes, unless precise
border delineation is required.
Artificial neural network techniques have been used in
remote sensing for feature extraction for classifying
multisource and multispectral image data. A characteristic
of such methods is that they may require a long training
time but relatively fast data classifiers, and unlike the
statistical classification methods, they are distribution-
free (Richards, 1993, p207, Beneiktsson and Sveinsson,
1997).
Information extraction in remote sensing can be either
directly or by inference by using a physical model of the
imaging process. However, the information which can be
extracted directly is limited. For land surfaces, it is only
possible to make the following measurements from
remotely sensed data: topography, albedo and
temperature. Inference method can potentially produce a
large number of measurements of the Earth surface and
atmosphere. However, unless the spatially and temporally
variable effects of the atmosphere are removed, the
feature extraction problem in remote sensing, if applying
theoretical models, is still intractable. The empirical and
semi-empirical approaches derived, based on the
knowledge and theory of the imaging process, either
traditional statistical methods or newly developed neural
network technique, remain the most attractive solutions
(Danson et al. , 1995).
3. TOWARD AN INTEGRATED APPROACH
3.1 The Needs for an Integrated Approach
SPOT panchromatic 10 meter resolution and
multispectral 20 meter resolution image data have to date
been the best satellite imagery data for civil applications.
International Archives ot Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
Within a short period of time, high resolution satellite
data of 1 meter resolution in panchromatic mode and 4
meters in multispectral mode will become available and
will provide potential applications for the user (Aplin et
al., 1997). The high resolution systems will provide an
order of magnitude improvement in ground resolution, at
the expense of less area and multispectral capability. The
investors in the systems believe that one-meter resolution
allows applications that require identifying, measuring
and mapping built objects such as roads and buildings,
and the real-time or near real-time availability of data in
digital form will make them immediately usable in GIS
database (Stoney and Hughes, 1998, Fritz, 1995). Many
photogrammetrists are skeptical of what the high spatial
satellite remote sensing systems can do for them,
compared to the proven high quality products delivered
by the traditional photogrammetric principles applied to
sub-orbital aerial photography. While for GIS users, they
are only concerned that the high spatial resolution raster
image data are spatially accurate, available in a timely
fashion, be reasonably priced, and integrate seamlessly
with their vector-based Geographical Information
Systems (Jensen, 1995).
Even though, the high spatial resolution image systems
will provide image data with less multispectral capability
when compared to some systems of coarser resolution,
they can be fused with lower spatial resolution
multispectral images by using band sharpening
techniques. In band sharpening, the product has the
spatial resolution of the panchromatic image and the
spectral characteristics of the multispectral image. The
spectral characteristics are useful for identifying thematic
features such as trees, water, soil, etc. With increased
spatial resolution, the features can be more accurately
delineated, thus making the resulting product more useful
for various applications, and even more useful if there is
no change in the spectral content of the sharpened
product. More importantly, band sharpening with a single
high-resolution panchromatic image allows the
multispectral band data to be acquired at a lower spatial
resolution. This permits systems to be designed that have
lower bandwidths and storage requirements. Lower
multispectral spatial resolution can also lead to the
implementation of increased spectral resolution on future
sensors (Vrabel, 1996).
The major purpose of feature extraction is to
automatically acquire information for mapping and GIS
databases. Accuracy, either geometric or thematic, is
very important and it must be similar or better than that
obtained by using manual methods (Trinder and Sowmya,
1997). With the availability of digital image data of high
resolution in a multispectral mode, region extraction and
edge finding must be presented complementarily. It is
important that researchers in the field of machine vision,
photogrammetry and remote sensing continue to
collaborate, so that advantages are gained from the
combination of the skills (Trinder and Sowmya, 1997). In
the next section, we propose a conceptual integrated
model for feature extraction by combining the knowledge
from different fields.
247
