of life. In
sed data it
en different
brightness
d unsealed
ur different
d and non
were used,
des
‘two surface
ratures can-
rs but ther-
or both sur-
C, solid line:
If some regions of interest have already been defined (in the
so called training of the data set or by an initial unsuper-
vised classification) it may be useful to take the distribition
of temperature values into account for classification purposes
(Holbe & Luvall 1989). We will show one possibility to use
this source of information.
We derive characteristic features of vegetational and non-
vegetational surface types from remotely sensed temperature
data which allow a grouping of the thermal image data into
these two categories. In order to assess classification accuracy
we compare our results with NDVI based vegetation detec-
tion.
2 DATA SET
The analyzed imagery was recorded by a DAEDALUS AADS
1268 line scanner on board a Dornier Do 228 aircraft yielding
a multispectral data set consisting of ten reflective bands
and one band in the thermal infrared. The experiment took
place at the 26 August, 1997, in late summer. We chose
parts of the city of Nuremberg, Germany, as test site. The
flight altitudes of 300 m and 900 m yielded geometric ground
resolutions of about one and three meters, respectively. The
data was recorded at 4:30, 8:30, 11:30 and 12:30 (MET),
in order to measure a characteristic part of the diurnal
temperature curve.
Atmospheric corrections were performed by using the SEN-
SAT/MODTRAN code (Richter 1992) relying on standard
atmospheric parameters supplemented by ground measure-
ments of meteorological data. In this paper we only use
brightness temperatures, because no correction for emissiv-
ities is possible (and required) to perform an unsupervised
classification. Having this simplification in mind, we use the
notion of temperature for abbreviation purpose. From the
data set a small scene showing the airport of Nuremberg
was extracted for every overflight. The early morning im-
age was taken as reference on which the images were reg-
istered using Akima local quintic polynomial interpolation
(Wiemker 1996). This scene is well suited for our work be-
cause it contains different types of concrete and roofs, dense
and light grass areas and trees. The four-dimensional tem-
perature data set resulting from the four overflights provides
the basis for the investigations described below.
3 THERMAL PROPERTIES OF DIFFERENT
SURFACE TYPES
In this section we briefly describe two kinds of information
contained in the data set that was used for classification.
3.1 TEMPERATURE CURVES
To take a first look at the data set described in Section 2
we defined several regions of interest from aerial photographs
representing different surface types. The corresponding av-
eraged temperatures of each region are plotted in Figure 2.
Grassland and concrete are represented by two curves each to
describe a brighter and darker region in the aerial photographs
respectively (corresponding to a brighter resp. darker col-
ored concrete and a thicker resp. more sparsely vegetated
grassland). For illustration purpose the two-dimensional his-
60 Vd. Ar V Rr TETE TT
[6] |
[5] |
[5]
= 407 7
Em [4] |
Us
S [2] |
o
s |
2 20} i
Ob que NN
4 6 8 10 12 14
Time (MET)
Figure 2: Temperature curves averaged over the regions of
various surface types ([1]: trees, [2]: grass (dense), [3]: grass
(sparse), [4]: concrete (bright), [5]: concrete (dark), [6]: various
roofs)
tograms of early morning and noon temperatures of the same
regions are shown in Figure 3.
With regard to a classification application based solely upon
the temperature data Figure 2 shows some remarkable fea-
tures:
e as expected from model runs trees and thick vegetated
grassland have flatter temperature curves with rela-
tively low noon values corresponding to high thermal
inertia values. It should be easy to separate these sub-
stances from other materials,
e roofs have the steepest curve and highest noon tem-
peratures,
e bright concrete and sparse grassland have nearly the
same heating behaviour and are expected to be hardly
distinguishable.
So temperature curves alone will not provide enough infor-
mation for a proper separation of surface types.
3.2 TEMPERATURE DISTRIBUTIONS
A second source of information based on thermal measure-
ments are the temperature distributions of various surface
types. On the right side of Figure 2 some examples corre-
sponding to the temperature curves on the left are shown.
It can be seen that roofs have temperature values of about
20° to 80° (of course several roofs were put together in
this region of interest) and so have the broadest temperature
histograms while concrete surfaces have the smallest.
To draw surface specific information from thermal infrared
images (Holbe & Luvall 1989) have modeled surface tempera-
ture distributions of forest landscapes using a two-dimensional
beta probability distribution. For our purposes the standard
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 229
