ıhe.de 
sensing 
ig-term 
r target 
a joint 
sensing 
rection 
cts by 
rection 
s of the 
by the 
s in the 
all land 
nt land 
  
  
neier & 
nethods 
iderable 
1alytical 
Cihlar, 
Ni & Li, 
er et al., 
tion of 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B-YF. Istanbul 2004 
Because partial cloud cover is practically always existent on a 
regional scale, different algorithms for creation of cloud-free 
AVHRR composites were developed. The composites are 
typically made from 10 day periods. 
BRDF effects in AVHRR composite imagery can visually be 
noted through a non-coherent and coarse structure as well as 
striping effects. The normalization to a standard viewing 
geometry potentially reduces significantly such effects. In the 
case of correlation of different indices with field measured plant 
water contents it is assumed that BRDF corrected AVHRR 
imagery will yield better correlations and thus enhance the 
certainty of predictions. 
2. METHODOLOGY 
2.1 Data sets 
In this study, AVHRR data from NOAA 15 and NOAA 16 
covering part or the whole Iberian Peninsula were used. The 
daily NOAA 16 data acquisition time is between 3 and 4:30 pm, 
data from NOAA 15 are usually received between 9:30 and 
11:00 am. The data is received in the HRPT format through a 
parabolic antenna. The data are calibrated as described by the 
NOAA manuals (KLM Users Guide, 2000). In order to avoid 
strong geometric and radiometric distortions, only data with a 
scan angle smaller than 45° were used. The image data are 
georeferenced and reprojected. The size of the imagery is 1136 
by 889 pixels. The data sets used in this study were received 
between May and September 2002. 
2.2 Models 
The model applied for normalization of AVHRR data in this 
study is the NTAM (Non-linear Temporal Angular Model) as 
described by Latifovic et al. (2003). It is based on the widely 
used Roujean model (Roujean et al., 1992). The Roujean model 
and the empirical modified Walthall model (MWM, Walthall et 
al. 1985) are used for comparison of model performance. The 
three models are defined as follows: 
The Roujean model: 
9(8,,0,,9) ^ ky t Kf (6,0. 0) * k, f,(8,,0,, 9) 
The modified Walthall model: 
0(8,0,,0) - ay (8, 8^) a/0, * a, cos(9) ^ a, 
The NTAM: 
248i) 
T % 
0,8.0,.6.0;) =|1+a,; (De 
L+(a, (+a, (NA-a) +a, (HA -a)*) £,(6,.6..0) 
t (a, () +a, (Ne +a, (Na) f,(6,,6,,9) 
where: 
(020.19) = d [x — 6)cosó ^ sino]tane, tane, 
ag (tan6, -- tan0, + - tan“ 6, - tan* 0, — 2tan6, tan, cos) 
n 
4] 
Op 
f.(6,.8,. 
A 9) 3m cos8,-cos6, 
| - eost sing |- 
cos = cos6, cos6, + sin6; sin6, cos à 
dos NIR - VIS - NDVI 
NIR + VIS 
a, =NIR - VIS 
and: 
p = reflectance 
¢ = relative azimuth angle = 10; - ¢, 
j = land cover class 
i = AVHRR channel (1 = VIS, 2 = NIR) 
The Roujean model belongs to the important group of 
physically based BRDF models that have reflectance as a linear 
function of parameters (Shepherd & Dymond, 2000). It 
considers the observed directional reflectance as the sum of a 
geometric and a volume scattering component. The parameters 
ko, kı and k, are semi-empirical coefficients representing 
physical properties of the surface. They are often referred to as 
‘kernels’ (Wanner et al., 1995). In model inversion ky, k; and k, 
are obtained on a per-pixel base through least squares fit 
between model and observations by minimizing an error 
function. 
Because the BRDF is time-wise dynamic, derivation of the 
model parameters in multitemporal data sets requires regression 
techniques made in subperiods (Leroy & Roujean, 1994). The 
number of cloud free observations within a sub period of about 
10 days can thus easily become a limiting factor (Cihlar et al., 
2002). The Modified Walthall model is a purely empirical 
model employing the four parameters a, a;, a, and as. 
The NTAM was designed at the Canadian Centre of Remote 
Sensing (CCRS) for normalization of AVHRR imagery from 
northern ecosystems (Chen & Cihlar, 1997; Cihlar et al., 2002). 
The defined aim of this study is to investigate its applicability in 
a Mediterranean environment. The NTAM modifies the 
Roujean model to a time-independent, non-linear, physically 
based 8 parameter BRDF model. Model inversion is to be 
applied on land cover classes rather than on a per-pixel base. 
The temporal dimensions of the NTAM are approximated by 
polynomials that are related to vegetation indices. They account 
for the varying green leaf area during the growing season and 
the land cover dependent patterns of geometric and volume 
scattering components (Latifovic et al., 2003). 
2.3 Cloud Mask Processing 
For the generation of cloud free input sampling data a cloud 
detection algorithm based on the formulations of Saunders & 
Kriebel (1988) is applied. The algorithm consists of five tests 
that are applied to each individual pixel. A pixel is only 
considered to be cloud free if all the tests prove negative. 
2.4 Spatial sampling 
The NTAM requires a database that assigns each pixel in the 
imagery to a land cover class. In this study, CORINE Land 
Cover data from Spain are used to assign land cover class 
membership. CORINE Land Cover describes land cover based