326 
the relations will not give the same result as the application of the original relationships to the averaged 
values of the parameters. The problems associated to the scaling of mathematical relations established from 
measurements at ground level to relations between values resulting from low resolution satellite 
measurements, becomes more important as such relationships deviate from a simple linear approach. Even 
in this case, such relations turn out to be non-linear at the low resolution level, if the determination of the 
independent variable in the relation can not be described as a simple linear average. 
The up-scaling problem from ground measurements to low resolution satellite data was approached 
by the Remote Sensing Group of the University of Valencia by using high resolution LANDSAT data in the 
transition, and by the integration between LANDSAT and AVHRR data for the analysis of the change of 
scale in the observation from satellites (Moreno and Melia, 1992). Integration of both data allows one to 
understand the meaning of low resolution measurements in terms of the spatial variability derived from high 
resolution data. The problem of defining the within-pixel variance at the AVHRR pixel scale has also been 
approached. Averaged pixel-values are derived from spatially degraded high resolution data, and the 
corresponding within-pixel variance maps are also derived from high resolution data. 
Specifically, the connection between ground measurements and low resolution satellite data 
(AVHRR) has been addressed by following a three-step procedure: 
(a) Accurate geometric location of ground measurements over the satellite image. In order to make it 
possible to locate in the image the points where ground data were acquired, an accurate method of 
navigation of AVHRR data is required. The geometric processing algorithms used (Moreno and Melia, 
1993a) has proved its efficiency and accuracy and it is possible to derive outputs of geometrically corrected 
data up to a pixel size of 10 meters, even for AVHRR. 
(b) Appropriate interpolation procedure, adapted to the viewing geometry, to determine the value of 
the corresponding satellite measurement from the pixels really measured. Since a ground point does not 
exactly correspond to a pixel in the image, we have to interpolate into a matrix of surrounding pixels to be 
able to determine the value for a new one not actually measured by the satellite. Standard interpolation 
procedures usually fail when accounting for changes in the overlap between adjacent pixels and changes in 
the pixel size and shape with viewing angle. Appropriate interpolation procedures, adapted to the viewing 
geometry of the AVHRR instrument, have been considered to be able to assign a value to the corresponding 
point in the image (Moreno and Melia, 1993b). 
(c) "Error" estimation in the satellite measurement, based on the comparison of data derived from 
low resolution satellite measurements with other satellite measurements at higher spatial resolution, e.g., 
LANDSAT TM (Moreno and Melia, 1992). This has been achieved by simulating low resolution satellite 
measurements from high spatial resolution satellite data and interpreting low resolution measurements in 
terms of the within-pixel spatial variability derived from the latter. A multi-resolution integrated dataset is 
then required to validate measurements over an extended geographical area, as well as a realistic method of 
simulating low resolution measurements from high resolution data (Moreno et al., 1992a). 
The NOAA-LANDSAT coincidence on June 12, 1991, one of the Golden Days, has been used to test 
the methods developed for multisource data integration. The 5S radiative transfer code (Tanr 6 et al., 1986 
and 1990) has been used for atmospheric correction of LANSAT TM and NOAA AVHRR data. 
A special analysis of the significance of NDVI values obtained from ground reflectance and from 
AVHRR data has been carried out. Both the pixel-averaged value and the within-pixel variance are analyzed 
as varying with spatial scale. The problem of determining an NDVI value corresponding to a characteristic 
point has been addressed by considering the resulting errors estimated for each measurement as a function 
of the corresponding spatial scale in the observation (Fig. 8 ). LANDSAT TM data were used as an 
intermediate step in the transition from ground to low-resolution AVHRR data. Two main conclusions may 
be drawn: (i) For the absolute NDVI values derived for each particular scale, significant differences are 
obtained when NDVI approaches 0.5 but they are very small for NDVIs close to zero. A parameterization of 
NDVI associated to each particular scale has been derived for the Barrax area, (ii) For the within-pixel 
NDVI standard deviation and its relationship to averaged NDVI values derived for each scale, for the 
Barrax area, that standard deviation at AVHRR scale shows a range of values similar to those obtained for 
averaged NDVIs, surely due to the high heterogeneity of the area. Such spatial properties of the Barrax area 
result very appropriate for deriving estimates of the dependence of the within-pixel spatial variability on the 
spatial scale. Thus, for AVHRR scale, a quadratic relationship bewteen the within-pixel standard deviation 
and the integrated NDVI values has been derived for that area. 
The spatial-scale variability of NDVI increases the difficulties in determining derived values of 
biophysical parameters (biomass, LAI, etc.) through models that establish (non-linear) relationships 
between NDVI and those parameters. A particular analysis was carried out by considering logarithmic 
relationships between LAI and NDVI, and observing the changes in the coefficients of such relationship 
when the spatial scale varies. Differences arc really high when high resolution (LANDSAT) and low 
resolution (AVHRR) scales are compared. The simple extrapolation of LANDS AT-derived relationships to 
AVHRR estimates yields to errors of more than 50%, particularly in the case of very heterogeneous areas 
(Barrax). In this case, a large range of NDVI values arc integrated when deriving low resolution values by 
using AVHRR data.