The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part Bl. Beijing 2008
3.3 Geometric Correction
The EnMAP level 2a processor produces ortho-images applying
the technique of Direct Georeferencing. The line-of-sight model
uses on-board measurements of the star tracker systems and
inertial measurement units combined by Kalman filtering for
attitude determination, GPS (Global Positioning System)
measurements for orbit determination (position and velocity),
and sensor look direction vectors derived from laboratory
and/or in-flight geometric calibration. An improvement of the
line-of-sight model can be achieved by automatic extraction of
ground control points (GCP) using image matching techniques
with reference images of superior geometric quality. Terrain
displacements are taken into account by global digital elevation
model (DEM) fused from different DEM data sets using quality
layers. Figure 4. illustrates this part (“Processing Chain, Level
2a” of Figure 2.) of the processing chain.
Figure 4. Geometric Correction
The accuracy of this rectification result is crucial for overlaying
the data with existing data sets, maps, or in geographic
information systems (GIS) and using them for evaluations like
change detection, map updating, and others. Therefore, first an
improvement of the line-of-sight vector with the help of
automatic extraction of GCPs by image matching is foreseen. In
order to automatically extract GCPs from the reference image a
hierarchical intensity based matching is performed (e.g., Lehner,
M. and Gill, R. S., 1992). The matching process uses a
resolution pyramid to cope with large image differences
between the reference and the coarse registered image. Based
on the Foerstner interest operator, pattern windows are selected
in one of the images and located with an accuracy of about one
pixel in the other image. This is done via the maximum of the
normalized correlation coefficients computed by sliding the
pattern area all over the search area. The search areas in the
matching partner image are determined by estimation of local
affine transformations based on already available tie points in
the neighborhood (normally from a coarser level of the image
pyramid). The approximate tie point coordinates are then
refined to sub-pixel accuracy by local least squares matching.
The number of points found and their final (sub-pixel) accuracy
achieved depend mainly on image similarity and decrease with
time gaps between imaging. Only points with high correlation
and quality figure are selected as tie points, including cross
checking by backward matching of all found points. Within the
next processing step the GCP information is used to estimate
improved parameters for the line-of-sight model by least
squares adjustment, including iterative blunder detection, which
eliminates step by step GCPs with a residual greater than a
threshold starting with the bottom quality GCP. This part of the
processor can only be used, if an appropriate reference image is
available.
The basis for all direct georeferencing formulas is the co
linearity concept, where the coordinates of an object point
r ™. t expressed in any Earth bound mapping coordinate frame
are related to image coordinates r sensor derived from the
measured pixel position in the sensor’s coordinate frame. The
rigorous relationship between 2D image coordinates and 3D
object coordinates is given by
r m =r m -t-V-R m .13 6oi ^ -sensor /|\
1 object 1 sensor ' “ body sensor 1 object ’ ' '
where li hody denotes the rotation from the sensor to the body
sensor
coordinate frame, which has to be calibrated, and R m denotes
body
the rotation from the body to a mapping coordinate frame,
which is derived from the angular measurements. If GCPs from
image matching are available, an additional boresight rotation
matrix can be estimated for refinement. The interior orientation
is described by mapping column values i to the sensor
coordinate frame with the focal length c by
N - 1R3: i r ™ = (x* ensor ,y™,-c) T . (2)
The scale factor s is determined by the intersection of the sensor
pointing direction with a given DEM using an iterative process,
which finally results in a 3D point in object space for each
image pixel. After object point reconstruction within the
mapping frame the coordinates are transformed to any desired
map projection, where the resampling (applying nearest
neighbor, bi-linear, or bi-cubic resampling) of the ortho-image
proceeds (e.g. Muller, R. et al., 2005; Muller, R. et al., 2007).
3.4 Atmospheric Correction
The EnMAP level 2b processor performs atmospheric
correction of the images employing separate algorithms for land
and water applications. Figure 5. illustrates this part
(“Processing Chain, Level 2b” of Figure 2.) of the processing
chain.
Processing Chain
I Processing Chain
Level 2b
Level 2b
(Atmospheric Correction)
(Atmospheric Correction)
Ovar Land
Over Water
Figure 5. Atmospheric Correction
The choice of the land and/or water mode is defined by the
customer. However, scenes may also be processed in both
modes, e.g. for coastal areas or inland lakes that may contain a
large percentage of land and water pixels.
Land Applications
Relevant criteria for the selection of a radiative transfer code
with respect to the EnMAP mission are:
• spectral coverage of the radiative transfer calculations
• spectral resolution
• aerosol models
• treatment of gas absorption and multiple scattering
The MODTRAN4 (moderate resolution atmospheric
transmission) code covers the solar reflective spectrum (from
400 nm to 2500 nm) and even the thermal region. It supports a
sufficiently high spectral resolution for the absorbing gases
(water vapor, ozone, oxygen, carbon dioxide etc.). It also
includes a rigorous treatment of the coupled scattering and
absorption processes. Moreover, it offers a set of representative
aerosol models (rural or continental, urban, maritime, desert).
Therefore, MODTRAN4 will be selected to compile a database
of atmospheric correction look-up tables with a high spectral
1268
