International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B-YF. Istanbul 2004
on a nomenclature of 44 classes that are hierarchically
organised into three levels (CORINE Land Cover Report,
2000). These data were reprojected to fit projection and cell
size of the AVHRR imagery.
The original land cover types were aggregated into 14 land
cover classes to represent major structural and surface cover
combinations. For each aggregated land cover class between 60
and 100 sampling points were defined. Minimum distance to
pixels of other classes was 5 pixels, minimizing thus the effect
of transition pixels. For these sampling points the median
reflectance value from a 3*3 window were sampled. Table |
lists the agglomerated CORINE Land Cover classes and their
relative fraction of the surface of Spain.
Aggregated | Comprises classes %
class coverage
12 12: Non irrigated arable land 23.08
13 13: Irrigated arable land 7:33
15: Vine yards
16: Fruit trees and berry plantations
17 17: Olive grows 3.13
20 18: Pastures 8.39
19: Annual with permanent crops
20: Complex cultivation patterns
21 21: Agriculture with natural 523
vegetation
22 22: Agro forestry areas 4.73
23 23: Broad leafed forests 6.88
24 24: Coniferous forests 9.00
25 25: Mixed forests 2.11
26 26: Natural grassland 5.54
27 27: Moors and heathland 3.54
28 28: Sclerophyllous vegetation 10.27
29 29: Transitional woodland shrub 6.45
31 31: Bare rock 2.42
32: Sparsely vegetated areas
0 All other classes (mainly 1.70
settlements)
Table 1: CORINE Land Cover classes and relative fraction of
surface
2.5 Temporal and angular sampling
Data were sampled from cloud mask processed datasets from
throughout the study period. Between 1000 and 2000
observations per aggregated land cover class were obtained.
Alternatively to cloud mask processing, data might have been
sampled from cloud free composites.
2.6 Model inversion and statistic parameters
For model inversion, the CCRS kindly provided its MIB
(Model Inversion of BRDF) software tool. For the linear
models, the MIB works on a matrix inversion based procedure.
For the NTAM, the modified Powell’s minimization method is
used to calculate the non-linear least squares fit in an iterative
procedure (Latifovic et al., 2003). Statistic parameters used in
the MIB are:
42
ae» . . 2 -
e The coefficient of determination R* defined by:
8°
2 = 1
R = M
where:
1 N
8° = N >b, m (acht
5? 1 N 2 1 N 2
obs = N > (04) E N 2. (aJ)
i=1 i=
e The standard error of the estimate (se) defined by:
2
Sb, = De]
where:
Po», = Observed reflectance
p = modelled reflectance
N = number of observations
3. RESULTS AND DISCUSSION
The CORINE Land Cover class 17 (olive grows), maybe the
most emblematic land cover for Spain, was chosen to
demonstrate the quality of the regression. In figure 2 modelled
and observed reflectance are displayed for the sampling data in
the NIR channel. 1052 observations were collected from NOAA
15 data and 852 observations originate from NOAA 16 data.
The data from the two sensors were inverted together. They are
displayed in different colours only for better legibility.
0,3 ;
[0] i
D 025 |
& |
S 0,2 =
© 0.15 + NOAA 15
© = NOAA 16
> 0,1 4
o
TS osse), 05 |
E |
0 ud
0 0,1 0,2 0,3 0,4
observed reflectance
Figure 2: Observed and modelled reflectances for olive grows
The statistic results of the inverse mode are very satisfying.
Figure 3 displays the coefficient of determination for the
different aggregated land cover classes. The mean coefficient of
determination for all land cover classes is 0.79 for the VIS and
0.83 for the NIR channel. Homogeneous land cover classes, e.g.
class 17: olive grows and class 24: coniferous forest, are
generally only slightly better modelled than more
heterogeneous classes. For the latter ones, the classes 12, arable
land, and 20, complex cultivation patterns may be named as
examples. This shows well how the NTAM accounts for the
varying amount of green leaf area within one vegetation class.