Limits and challenges of optical very-high-spatial-resolution satellite remote sensing for urban applications - Urban Remote Sensing: Monitoring, Synthesis and Modeling in the Urban Environment - page 34

(a) (b)

FIGURE 3.1 A sample of multispectral Quickbird data covering the downtown of San Francisco, USA: (a) the original image;
(b) the output after a supervised classification.

3.3 Spectral problems

differences in the apparent spectra. The effect is well-known in
hyperspectral data analysis (see for instance the work by Schiefer,
Hostert and Damm, 2005) but still to be considered for sensors
with a smaller set of wavelengths.

VHR images in urban areas may be used for many applica-
tions, some of them connected to environmental monitoring.
Mapping of dangerous materials (Marino et al ., 2000), ''urban
forest'' monitoring (Xiao et al , 1999) and sealed part detection
(Segl, Roessner and Heiden, 2000) are only examples. In these
applications, however, the stress on VHR resolution is due to
the need to detect the exact location of the above mentioned
features, which are of course at the level of single objects in urban
areas, and thus require a sub-meter spatial resolution. It is how-
ever extremely important for these applications the availability
of VHR imagery also with very high spectral resolution. This is
immediately obtained by observing that most of the above men-
tioned examples refer to airborne hyperspectral imagery, which
pair spectral and spatial fine resolution. The availability of spe-
cific spectral bands for urban roof material mapping have been
discussed for instance in Herold, Gardner and Roberts (2003)
showing that there are specific frequencies in the 2 m wavelength
range that are extremely useful to map urban materials.
Generally, the current operation VHR optical sensors from
satellite platform do not match the requirements in term of
spectral bands for many urban applications. Their use beyond
basic mapping is hampered by the lack of important bands in
the medium infrared, which can be used to characterize building
materials and surface properties (e.g., for road asphalt status
monitoring as in Herold and Roberts, 2005). WorldView-2 is
somehow the next move in this direction, as it couples VHR
spatial resolution in the panchromatic sensor with eight bands in
the multispectral one.
A different problem, which has not been addressed so far
in urban research, is the need of a complete characterization
of the bidirectional reflectance function for urban materials. In
fact, the dependence of the material spectral signature from the
orientation between the sensor and the observed specimen is an
issue when spatial and spectral fine resolution are paired. For
this kind of sensors, per-pixel analysis must take into account
the effect or the classification will results in errors because of

3.4 Mapping limits
and challenges

As a consequence of the challenges related to the fine spatial
resolution and the relatively coarse spectral resolution of current
optical VHR data from spaceborne platforms, urban mapping
is not an easy task. Although there are studies focused on the
trade-off between these two characteristics (comparing airborne
and satellite sensors, see Gamba and Dell'Acqua, 2006), they are
limited to the specific considered legend and still need to be
validated on large and various data sets.
As a result of the geometrical and spectral inconsistencies
of single pixels, the stress in urban mapping using optical VHR
sensor is on using information frommore than one pixel at a time.
Spatial aggregations of pixels are considered by means of textural
(Yashon and Tateishi, 2008 or geometrical features (Lisini et al .,
2005), or by tightening the class assignment for a pixel to the
class membership of its neighbors (e.g. by means of a Markov
Random Field approach, as in Gamba and Trianni, 2007). The
most exploited technique, however, is the so called object-based
image analysis (OBIA, see Blaschke, Lang, and Hay, 2008). The
idea is that objects (i.e. portions of the scene that are statistically
consistent with respect to a set of spectral and spatial features)
are to be searched for. They define a higher level of interpretation
of the scene beyond the land cover obtainable from the single
pixel. By aggregating the features at the pixel level it is possible
to assign each of these objects to a class, with either a soft or a
hard decision. Eventually, both the pixel-based and the segment-
based decisions can be profitably used together to characterize
the scene.