04
an
of
er
ke
ral
ler
on
mn.
1€-
nd
)0;
to
the
ols
we
om
the
are
the
ixel
atic
nal
ata,
SY
his
the
"ing
rate
| on
are
grey
area
ulti-
olds
cture
)04).
rast,
erns.
ean,
dow.
TCy
C
ga
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
values in an image and analyse the grey level pattern and
variations in a pixel's neighbourhood by determining pixels
positions that have equal or nearly equal grey values. This grey
level variation can be directional or not. While the mean and
variance provide a simple description of the statistics of the grey
values, the entropy measure provides spatial information of the
grey values related to their directionality and frequency of
occurrence. Thus it can be used to detect and extract image
regions based on the relative degree of randomness of their
structure patterns.
When the feature is linear or an edge in an image, edge
detection methods can be used to determine sharp changes in
the pixel values. These changes in brightness in the two-
dimensional image function, /(x,y), are determined by various
edge-detection operators based on the two directional partial
derivatives, (d//dx, dl/dy), which are approximated as image
pixel differences. Mapping feature operations required selection
of specific edges rather than all edges in the image and as much
as possible low error in edge detection location and type. The
most commonly edge detection operators are the Sobel, Prewitt
and the Laplacian. Their disadvantage is that they are sensitive
to noise and they might produce more than one response to a
single edge. Therefore, one edge operator that is recommended
for mapping feature extraction is the Canny operator (Canny,
1986). The Canny operator produces a low error in the
detection of an edge, keeps the distance between the detected
edge and the true edge to minimum, and has only one response
to a single edge (El-Hakim, 1996).
Finally, the quality of feature extraction can be improved with
the integration in the process of existing knowledge either as
part of the process or as additional constraints. For the former,
better knowledge of the type of the training areas will result in
high classification accuracies and therefore to higher extraction
accuracies. For the latter the idea is based on the principle of
determining and establishing conditions that uniquely
characterized the features of interest in order to increase the
success of recognizing and extracting these particular features
from image. These conditions can be applied as *pseudo" bands
such as a DEM layer, which can be included in the
classification process to improve the classification results for
extracting vegetation or buildings (Eiumnoh and Shrestha,
1997; Hodgson et al., 2003). Or they can be applied as spatial
constraints, where the extraction of a feature is based on the
intersection of conditions-derived spatial layers using logical
operators.
Currently most of the efforts for automated or rather semi-
automated feature extraction are concentrated on thematic type
of extraction, such as roads, water bodies, vegetation, and
buildings (Auclair et al., 2001; Jodouin et al., 2003; Baltsavias,
2004; Zhang, 2004).
2.3 Change detection
Change detection requires the comparison of two temporal
datasets for the identification and location of differences in their
patterns. Although in many cases the comparison must be
conducted between heterogeneous datasets, for example “new”
image and “old” vector database data, the actual comparison is
conducted with homogeneous types of data. That is, the change
detection is reduced between image data or between vector data.
The former is referred as image-to-image change detection, .
while the later as feature-based change detection.
613
2.3.1 Image-to-image. In the case of multi-temporal images
we can distinguish two basic approaches. An indirect image
change detection, where the change analysis follows an image
classification process. The comparison can be done by either
differencing the two raster classified thematic layers or by
extracting the boundaries of the thematic regions and conduct a
vector (i.e., feature-based) change analysis. With this approach
we overcome problems related to image acquisition conditions,
such as different sensors, atmospheric and illumination
conditions and viewing geometries. The accuracy of the
detected changes is proportional to the accuracy of the image
orthorectification and of the classification results.
The second approach is the direct comparison of two temporal
images (Singh, 1989). Various techniques supported by the
functionality of IP and GIS systems are:
image differencing, where the two co-registered temporal
images are subtracted pixel-by-pixel. This approach is affected
by the various image acquisition conditions and some form of
radiometric normalization is applied to both images to reduce
these effects. Still the determination of the threshold between
change and no-change in the histogram of the difference image
is a critical issue for the resulting changes.
* image ratioing, where the ratio of the values of corresponding
pixels between the two temporal images are computed. If there
is no or minimal change the ratio is close to 1. Again some form
of radiometric normalization between the two images needs to
be applied, while the selection of the threshold is critical as
well.
* image regression, where the pixel values of the second image
are assumed to be a linear functions of the corresponding pixel
values of the first image. A least squares regression can be used
to determine the linear function. Using this function the
estimated pixel values for the second image can be computed.
The difference image is determined between the estimated
second image and the first image using either image
differencing or image ratioing. If there is no change the pixel
values of the unchanged areas will be close to the estimated
pixel values, otherwise there will be changes.
* principal component analysis (PCA) for multispectral
multitemporal images, which can be applied either to each of
the images and the principal component of each data can be
compared with one of the above methods, or can be applied to a
combined image consisting of the combined bands of the
images to be compared.
2.3.2 Feature-based. For the feature-based approach various
functions of spatial analysis are used, such as layer union, layer
intersection, buffer generation, and topological overlay. The
spatial change AS; » is defined as the difference between the
spatial union of the two temporal homogeneous vector datasets
S; and 55 minus their common spatial elements (Armenakis et
al., 2003):
A812 7 (Sy 085 - (Sj ^85)
-(Sj- (Sj 82) 0 ($5- (Sij 85)
= Del 9 Agg