The International Archives of the Photogrammetry, Remotó Sensing and Spatial Information Sciences. Voi. XXXVII. Part B7. Beijing 2008 
However, the spectral range of the MFR instrument prohibits 
the calculation of E dir for wavelengths above lOOOnm. To 
overcome this limitation, the utilisation of MODTRAN data is a 
possible option to estimate the direct irradiance up to 2500nm. 
3.1.4 Saturation Detection 
The detection and filtering of saturated, upward looking 
measurements (L inc di ffj c ) taken around the sun direction is 
necessary due to (a) incorrect radiance values of these readings 
due to saturation and (b) interference with the BRF retrieval 
scheme that requires E dir and L inc di ff to be measured as separate 
entities. 
The filtering thus removes potential direct irradiance 
components from the upward looking (L incdi f f j c ) dataset. 
Figure 3: FIGOS Pre-Processing Dataflow Diagram 
3.1.5 Temporal Correction 
FIGOS data acquisition takes around 20 minutes per 
hemisphere. During this time the illumination angles and 
atmospheric conditions are changing. Consequently, the 
radiance incident upon the target is not constant and the 
recorded reflected energy is time dependant. This hinders the 
BRF retrieval as the diffuse, angular irradiance L mc di ff_ sd is not 
available for the whole hemisphere for every point in time a 
target sample is taken. The temporal correction tries to account 
for changes in the diffuse irradiance by applying ratios derived 
from the E dlff MFR . 
3.1.6 Hotspot Shading 
The FIGOS instrument has been built to minimise the hotspot 
shading by eccentric placement of the zenith arc (Schopfer et al., 
2007a). The used fore optic is thus the only part that can 
produce shading effects (~lcm in diameter). Potentially 
affected readings can be selected based on the geometry and 
either be omitted from processing or replaced by interpolated 
values using the neighbouring samples. 
3.2 Optional Pre-Processing Operations 
Once the quantities E din L r and L inc di ff have been calculated as 
outlined in the previous section, further subsequent operations 
might have to be applied. These may include: (a) removal of 
noisy bands, e.g. the well-known water vapour absorption 
wavelengths, (b) spectral dimensionality reduction, e.g. 
downsampling or (c) sensor convolution. The latter two 
transformations can both drastically reduce the volume of the 
data in the spectral domain and could increase the speed of the 
BRF retrieval considerably. 
As the application, configuration and sequence of such 
transformations is not known a priori, the system must offer 
freely configurable, user definable processing (Hueni & Tuohy, 
2006). 
3.3 Generic Considerations 
Processing operations are applied to data in a certain order and 
the according dataflow can be described by a directed graph, as 
can be observed in Figure 3. Such a network consists of 
processing modules and data sources/sinks. In order to avoid 
code redundancy and provide the flexibility as defined in 3.2, a 
generic approach is required that allows the application of 
modules to spectral data originating from the SPECCHIO 
database. 
A further important requirement is the independence from 
specific sensors, i.e. processing modules should be applicable to 
data stemming from differing instruments in terms of number of 
bands, central wavelengths and spectral response functions. 
3.4 Concept of the Space Processing Chain 
A solution to the generic, flexible requirements outlined above 
is given by the concept of the ‘Space Processing Chain’. It is 
based upon the definition of spectral spaces and feature spaces 
by Landgrebe (1997). The continuous, spectral response of an 
object is transformed into a discrete space by the sampling 
instrument. This transformation is defined by the sensor 
characteristics (Hiini et al., 2007a). Thus, data captured by 
different instruments ard contained by different spaces. A space 
has a dimensionality equal to the number of bands of the used 
sensor and the spectra are data vectors contained within this 
space. 
Processing modules are effecting a transformation on a space, 
i.e. the spectral data vectors of the input space are transformed 
to an output space. The algorithm of the processing module 
defines the dimensionality of the resulting space. This is 
illustrated in Figure 4 with an input space of dimensionality N 
being transformed into another discrete space of dimensionality 
M. 
Figure 4: Transformation into a new space by a processing 
module 
The Space Chain concept has been implemented in Java as part 
of the SPECCHIO application and offers graphical 
representations of the spaces and processes, thus allowing the 
user to check the effect of operations on the resulting spaces 
interactively.