2.3.2 Spectral libraries and classification: Once we get the
multi-index image, and through a process of image
interpretation, we selected the categories of land cover, also
taking advantage by the use of Google Hearth platform, which
provides higher resolution images, in order to corroborate the
categories chosen. For defining land cover classes we
proceeded as follow: first of all, three possible types of urban
areas were selected; afterwards the potential industrial
categories, and then plenty of possible kinds of terrain. After
that, it has been selected the green surfaces, and the waters
depending on the degree of depth. Tools of automatic
vegetation delineation and relative water depth detection were
used to better identify the above categories. The surfaces were
selected through the instrument of ROI, and for each ROI
selected it was calculated class statistics. The mean value, for
each class and at each band (previously normalized in the 0-255
scale, as mentioned before), has been used to generate a
spectral library with a final set of 48 land cover classes (Figure
3-4). This procedure is also important to analyze the specific
characteristics of the materials, depending on the indices, which
is useful to make corrections at the curves to improve the
following classification based on the spectral library.
170
wo
Brightness
SLOPE
Ol
VariaecelL$U21
Contrast SU21
FomopgereityLSU21
ë
INDICES LIBRARY
— Urban 1 — Urban 2 Streets ——industrial l industrial 2
Industrial Industrial à —SOt 1 il2 ——50it 3
Sole EE Soll 8
GT . c7 w-5MD- URES Cee Soil 11 7750413
v Soil 12 vou SOI 16 v SOIL 18
s Stil EO v Seil 21 1422 St 23
ee Soil 22 CroplANDL c CropLAND2 7 EvengreenFORES11
v Evergreenf OREST2 FORESTI EyergreenFORESTA p. E DecicuausFORESTI
DeciduübusFOREST? dousFOREST3 DrwtG ShadedVEG Shallow
Moderate Deep Snow
Figure 3. The spectral library, based on mean values, used to
classify
255 VEGETATION ANALYSIS
— CropL ANDE
n CropLAND2
— fvergreenFORESTI
EvengroenFOREST2
co EvergreenFORESTS
Evergreen ORESTA
-SponteneocsVEG
DecitduosFORESTI
Dociduous FORESTA
Deciduous FOREST
= S DryVEG
ee : ShadedvEG
sun
tua
NDS
43
MSA
NOVA
MNOWI
iU
Band:
Bands
vu: WS
[CE]
Brightness
Greenness
Nani Si
Varin
Homogeneity Su 21
Figure 4. Example of the spectral library for vegetation classes
Automated pixel-based multispectral classification of our multi-
indices imagery is applied by using Minimum Distance
algorithm and the spectral library above mentioned. The
Minimum Distance technique calculates the Euclidean distance
between each pixel within the image and the average value, for
the specific class of land cover, represented in the spectral
library. The classification has been repeated for all Autonomous
Communities of Spanish Mediterranean coast always using the
same spectral library of figure 3, and we found out a really
good homogeneity of results in the comparison between the
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
different classifications of land cover for the Autonomous
Communities. Figure 5 shows an example of the results of
Minimum Distance classification and for two Autonomous
Communities: Cataluña and Valencia.
Figure 5. Land covers classification result by applying
Minimum Distance technique.
2.4 Interpretation of primary results, correction,
and accuracy analysis
Although the result of the classification provides a broad
overview about the class composition into the landscape under
consideration, we focused the analysis on the impervious land
cover classes. Actually a review of primary results, evidences
interesting outcomes in discriminating artificial surfaces and
natural land cover classes such as vegetation, water, cropland or
bare lands. The process also allows certain goodness in
identifying different impervious typologies, such as residential
and industrial settlements, even it is mixing with some
categories of terrains. A second step of classification was
applied in order to clean the mistakes due to upon mentioned
mixture. This classification is based on parallelepiped algorithm
and just aims to reclassify those terrains which mixed up with
urban functions. Parallelepiped classification uses a simple
decision rule to classify multispectral data. The decision
boundaries form an n-dimensional parallelepiped classification
in the image data space. The dimensions of the parallelepiped
classification are defined based upon a standard deviation
threshold from the mean of each selected class”. It was
reclassified four types of terrains, selected by using ROI tool,
and the result was overlapped on the primary classification to
improve the result (Figure 6). Red tone colours indicate
residential areas, while yellow and cyan show industrial estates.
* ENVI help guidelines
