INEAR
an area
ed with
atistical
lying a
portions
and to
omplex
ve all
notable
à kind
more a
ographs
by field
f N-W
1e main
cultural
es, the
. open
aa
eas.
tinuous
%), 1.2
3s), 2.1
bodies
coastal
2:345
astures
shrubs
ence of.
shrubs
sity of
forests
% and
greater
uercus
pubescens. density between 35% and 70%), 6.4 broad-leaved
high density forests (density greater than 70%), 6.5 coniferous
veforestation . 6.6 broad-leaved reforestation..
The definition of the endmembers and vegetation continuum
legend has been obtained starting from the nomenclature
described above: this nomenclature has been further refined and
integrated by iterative spectral analysis of the multitemporal
dataset. evaluation of the classification results and assessment
ofthe errors occurred in the mixing model analysis.
1.2 Image data
Two Landsat 5 - Thematic Mapper data quarter scenes (track
193. frame 92) were acquired to obtain spectral and
multitemporal discrimination of the vegetation canopies: in
order to get discrimination capability in evergreen and
deciduous quercus species. as well as for pasture classes and
shrubs, a mid spring date and a deep summer date were chosen:
31st March and 22nd August 1994.
About half of the quarter scene (80Km X 40Km) was processed
to obtain the land use and vegetation map. About 20% of this
area resulted recently burnt in the August image. reporting
highest fire events occurred at the beginning of that month.
2. METHODS
2.1 Geometric and radiometric preprocessing
A first geometric rectification process was made in order to
correlate the two Landsat scenes each other; due to the almost
identical flight direction of the satellite overpasses the
registration was performed with a simple 1st order polynomial
transformation with a rotational coefficient very close to 0.
The projection on the UTM 32 coordinate system was made
after the classification process in order to limit the resampling
effects on radiometry and on the multitemporal registration.
During this last geometric rectification process no digital
elevation model correction was applied due to the smooth
landscape and low elevation range occurring in the study area;
the estimated maximum planimetric error introduced by this
approximation for such an elevation range is less than 20 mt,
which is comparable to the planimetric location error obtained
in the control points and test points used for the UTM
projection transformation. This error is compatible with
1:100000 scale products geometric specifications.
Despite the low elevation range. the typical Mediterranean
landscape is often characterized by steep slopes and
discontinuities that create small but serious topographic effects
that cannot be recovered easily without an accurate DEM
(generally unavailable or extremely expensive) and a
comparable spatial resolution in the imagery. To recover these
effects spectral normalization has been applied to the two
Landsat scenes by means of chromatic components computed
Starting from each TM channel reflectance.
The chromatic transformation, while giving generally good
results in topographic effects normalization, maintains the
linear relation originally existing among the total pixel
reflectance and the sum of area-weighted reflectances for each
endmember class. thus conserving the necessary hypothesis to
allow the application of linear unmixing methods.
575
2.2 Training sets, endmembers and channel selection
Within the study area, a square of 500 x 500 pixels was
selected. This area is representative of the most important
cover classes we put in the described legend. With the help of
aerial phothograps, showing the same zone, a reasonable set of
polygons were selected for each classes on a RGB 432
composite of the Thematic Mapper august scene. In this way, a
overall number of 73 polygons was defined. The standard
statistical parameters, referred to each training-set, were then
extracted from all original and all chromatic bands. Using
different kinds of digital tools. e.g. scatterograms and plots, it
was analyzed the spectral behavior of each cover classes within
the available spectral space. This spectral analysis allowed to
estimate the discrimination between the cover classes
considered in the legend and it was very useful for the choice of
the better set of bands to use in statistical classification. It was
observed a good spectral separation between the land use-land
cover classes hereinafter indicated.
The definition of endmembers, which represents a crucial point
in the unmixing method, has been driven starting from the
analysis of typical vegetation components, the ground truth plot
location on the imagery and the subsequent spectral distribution
analysis using tools available with EarthView. While running
this activity various interesting aspects arose among which the
presence of further endmembers and the usefulness of
visualization of spectral spaces in endmembers choice.
In our experiment, after this analysis, five classes were
indicated as endmembers: three species of woods (quercus ilex.
quercus pubescens, quercus suber), a typical shrub association
of secondary species (Cistus spp, Pinaccia Lentiscus) and an
open pasture class were chosen as endmembers, since they
represented the main vegetation species in this study area:
other typical vegetation species, even present in the Sardinia
district. (e.g. coastline maquis), were excluded for their
extremely scarce presence in the study area.
2.2 Land Use classification
The general land -use classification has been driven by the
following steps: a first raw classification allowed to separate
water bodies, bare rock outcrops and permanently bare soils by
multitemporal NDVI thresholding. Secondly, manual
identification of urban areas, recently burnt areas and
horticultural areas characterized by small-parcels was done.
Finally the Mahalanobis classifier was applied to classify the
remaining areas in the following other classes: quercus woods,
shrubs, pasture, 3 crop types and coniferous reforestation.
2.3 Linear unmixing classification
The classification of natural vegetation continuum has been
obtained using Linear Mixing Models. Linear Mixing Models
allow to classify multispectral satellite images at subpixel level.
For each pixel proportions between a number of different
materials are determined recalling that the spectral radiance
recorded by a satellite results from a mixture of the energy
reflected by all surface types within the instantaneous field of
view of the satellite sensor. The main problem with these
methods is a correct identification and characterization of a
reasonably small and representative number of surfaces whose
mixture can reproduce the spectral distribution observed. It's
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996