Image Processing Reference
In-Depth Information
mapping accuracy is high. Some surfaces indicate such
unique characteristics like green vegetation, red tile roofs,
or swimming pools. Other land cover types seem spec-
trally alike. In particular, darker built up surfaces like
asphalt roads and dark tile and composite shingle roofs
have quite similar signatures and no distinct absorption
characteristics. Considerable confusion between these
surface types is anticipated in remote sensing applications.
One remote sensing mapping approach to test this is shown in Fig.
4.3
. It presents
the results of a matched filter analysis (ENVI image processing software) applied
to hyperspectral AVIRIS data (4 m spatial resolution) acquired over Goleta,
California.
Matched filter analysis compares the spectrum in each image pixel to a known
reference spectrum from a spectral library. If the spectral signal of a specific pixel
and the reference spectra perfectly match, hence the reflectance properties of the
surfaces are the same, the matched filter score becomes 100; if they are absolutely
dissimilar the score is 0. The first frame in Fig.
4.3
shows the matched filter scores
for red tile roofs. It indicates that only a few areas have similar material characteris-
tics than the sample red tile roof spectra. The areas with higher matched filter scores
actually are red tile roofs. These roofs can be mapped with high accuracies from the
AVIRIS data. Wood shingle roofs also have fairly unique spectral signals and the
matched filter approach identifies these roofs with high accuracy. There are, how-
ever, a few areas in the left part of the image that show intermediate to higher
matched filter scores. The most significant false positives appear to be open fields
consisting of senesced grass in more rural areas. Both wood shingle roofs and
senesced grasslands contain non-photosynthetic vegetation and therefore their spec-
tral signals are somewhat similar (see also Fig.
4.2
). In contrast, the matched filter
for asphalt road shows considerable spectral confusion. In this map (lower frame),
the matched filter correctly identifies most road surfaces, but also maps large areas
of other surface types like composite shingle roofs, parking lots, and tar roofs. The
spectral similarity was already indicated in the spectral signatures of Fig.
4.2
. In fact,
these land cover types are composed of similar materials like dark asphalt, rocky
components, and other tar/oil products. This fact shows that we have to consider
whether we are interested in mapping urban materials or urban land cover types.
From a material perspective urban areas are composed of
four main components: (1) rocks and soils (i.e., minerals),
(2) vegetation, (3) oil products such as tar and asphalt, and
(4) other human-made materials such as refined oil products
like paints and plastics. Most surfaces represent an aggre-
gate of these components, i.e., concrete includes rock aggre-
gates, limestone, and portlandite. The matched filter
approach is best suited to identify and map urban materials
using the full spectral range. Other methods focus more on
small-scale spectral absorption features to derive detailed
material characteristics (e.g., mineral compositions,
spectral similari-
ties among some
of urban materials
pose challenge for
urban remote
sensing
application
urban surfaces are
generally made
of one or more of
four material
components:
minerals, vegeta-
tion, oil products,
and human-made/
artificial materials
