Environmental Engineering Reference
In-Depth Information
OBIA or spatiospectral analysis of the data at a single spatial
scale must however be complements with the ability to adapt
to data with different spatial resolutions and different viewing
angles (here called
spatial adaptiveness
). To this aim, most of
the proposed and more advanced methodologies are based on
adaptive combined spectral and spatial processing of the whole
scene. For instance, in Bruzzone and Carlin (2006) an initial
segmentation of the scene at multiple resolutions is performed,
to adapt the data interpretation procedure to the different scales
of the scene objects.
Moreover, interpretation algorithms may fully exploit the
available information in the scene by exploiting
class adaptiveness
.
The concept behind this term is that different classes require
different processing steps, or different parameters for the same
sequence of processing steps, and tailoring general algorithms to
these requirements (in addition to different scales) is a way to
furtherimprovetheireffectivenessforeachclass.
By combining properties at different scales, spatial character-
istics and class-tailored processing, mapping algorithm for VHR
data are able to go beyond basic land cover mapping. However,
to this aim, and especially in urban areas, they need to accom-
modate the
a-priori
information by the human interpreter or the
ancillary data that often are available for these areas. So far, one
of the still open and more interesting challenges of VHR optical
urban remote sensing is to define a strategy useful to mapping
land use with a consistent legend in many parts of the world. The
challenge is the development of these methodologies, robust to
different geographical locations, and able to capture the variable
spatial scales of land use classes in this environment, is one of
the most important to move from ''academic exercises'' to the
operational use of VHR optical data.
to exploit OBIA or feature based comparison to avoid these
problems. Object-based comparison was proposed by means of
the exploitation of a scene model by Hazel (2001) and by Car-
lotto (1997). In the latter paper non-linear prediction technique
for measuring changes between images and temporal segmen-
tation and filtering techniques for analyzing patterns of change
over time are used. Other examples of object-based change detec-
tion can be found in Dekker (2004), Walter (2004), Niemeyer
and Nussbaum (2006), Blaschke (2003), and Benz
et al
. (2004).
As a different approach, geometrical features may be exploited,
often in addition to pixel-based change maps, to detect impor-
tant elements of the scene (Lisini
et al
. 2005) and changes
(Dell'Acqua, Gamba and Lisini, 2006). No ultimate approach
has been designed so far, but many
ad hoc
solutions for specific
problems have been devised.
3.6
A possible way forward
Summarizing the analysis of the previous sections, it is expected
that a possible way forward, able to design efficient and effective
ways to exploit VHR optical data in urban areas, is to match
spatially-aware and adaptive classifiers to object-based analysis
and change detection. In the following we will show one possible
way to implement these ideas together with some results to
validate it.
We will start first by the above mentioned notion of
class
adaptiveness
to design an example of such a scene classifiers. As
discussed in the introduction, spatial feature are often peculiar
to specific classes, especially in anthropogenic environments.
Furthermore, this idea could be widened to spatial relationships
among elements of the same class or of different classes in the
same environment. Therefore, the idea is to exploit these char-
acteristics by means of a class-specific spatial post-processing of
a basic pixel-based classification map. The key is to define and
apply a set of rules in a different and adaptive way to each of
classes of the original classification map, as depicted in Fig. 3.3.
This set needs to be designed including flexible and efficient
spatial operators, and must able to reflect the known information
about the spatial features of the objects belonging to a class. The
operators need to be designed to reduce misclassifications and
discard wrong associations between scene objects and mapping
classes due to errors in the first classification step. According
to this description, the best set that we could think of is com-
posed by some of the operators of the so called ''mathematical
morphology''.
Mathematical morphology is a theory for the analysis of spa-
tial structures, since it aims at analyzing the shape and form
of objects. The most important operations can be grouped into
morphological and geodesic transformations, and are defined
on single-band grey level image
X
3.5
Adding the time factor:
VHR and change detection
One additional set of challenges and problems of VHR optical
images is related to the exploitation of the time factor, i.e. the use
of multitemporal data. In order to find changes or to highlight
persistent features in coarse and high resolution optical images
it is often enough to require the same viewing angle (which for
most sensors is fixed) and ascending or descending orbit. This
is no more the case for VHR images, which might be affected
by (co-)registration errors due to the uncertainties in DEMS or
orbital measures, as discussed in the previous sections. There is
therefore the need for methodologies meant to cope with the so
called ''registration noise'' and provide robust change detection
at the pixel level. As an example, we refer here to Bruzzone and
Cossu (2003) and Bovolo, Bruzzone and Marchesi (2008), both
very effective approaches. Additionally, as an example of the
problems researchers are called to face, Fig. 3.2 shows pre- and
post-earthquake images taken over the town of L'Aquila, central
Italy. The earthquake struck the town inApril 2009, causing death
and damages. A pixel-by-pixel post-classification comparison of
the center of the town in both images is shown in the figure. It is
clearly visible that most of the changes highlighted between the
two classifications are simply due to coregistration errors, while
the real changes are ''lost'' among the false positives.
The trend in multitemporal urban remote sensing using VHR
optical data is, as already mentioned for single date classification,
X(i,j)
.Formulti-spectral
images, operators are computed separately for each component
and then combined to obtain a scalar value. Usually the max-
imum of the magnitude among the bands is considered. In
the following, the main morphological transformations will be
quickly introduced. A longer and more precise description can
be found in Soille (1999).
The main morphological transformations are:
=
Erosion. The erosion of a set
X
by a structuring element
B
is
defined as the locus of points
x
such that
B
is included in
X
●
Search WWH ::
Custom Search