In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
344 
Figure 2. Footprints of 3 km strips 0818, 0825, and 0833. 
2.3 Geometric and radiometric post processing of the 
ADS40 data 
Direct georeferencing was used in aerial triangulation with 59 
control points and sub-pixel accuracy was reached. Leica XPro 
(4.2) was used for producing three radiometrically corrected 
versions of each MS image: 
• at-sensor radiance data (“calibrated” option): ASR 
• atmospherically corrected target reflectance data 
(“atmospheric” option): ATM 
• atmospherically and BRDF-corrected data 
(“atmospheric + BRDF” option): FULL 
All three were produced for the 2-4 km strips, while ASR and 
ATM for the 1 km data. The atmospheric correction and reflec 
tance calibration in XPro is based on the radiative transfer eq 
uation by Fraser et al. (1992). This atmospheric correction re 
sults in images, where the digital numbers are calibrated to 
ground reflectance. BRDF correction is based on a modified 
Walthall model. The details of the correction methods are pre 
sented in (Beisl et al., 2008). All corrections rely on a priori 
camera calibration and parameters derived from the image data. 
We used the default software settings in XPro processing. 
2.4 Radiometric in-situ measurements and quality assess 
ment of the ADS40 images 
Ground measurements of reflectance targets were carried out 
during the overflight. Targets included reflectance tarps with 
5%, 20%, 30%, and 50% nominal reflectance and well-defined 
surfaces (fine sand, grass, asphalt, gravel, hay). The ATM 
images were validated by Markelin et al. (2010) for the nadir 
reflectance. We sampled the targets by 4 x 4 -m rectangles and 
pixel data were analyzed for precision. 
2.5 Photogrammetric operations in the ADS40 data 
We implemented an ADS40 sensor model into the digital pho 
togrammetric workstation KUVAMITT, guided by source-code 
samples from Leica. All analyses were done in epipolar images, 
where the distortions due to the camera movements are remo 
ved. In ADS40, the exterior orientation parameters are needed 
for each scanline. These were defined in a local XYZ system, 
which had a 3D offset and rotation with respect to the WGS84. 
We used accurate transformations to reach the coordinate 
system of the trees. Each CCD line had the xy(z) camera coor 
dinates of the 12 000 pixels. The mapping from 3D to image 
was solved by iteration that limits a range of scanlines for a 
final sequential search of the pixel position. We used a nearest 
pixel interpolation. 
2.6 Extraction of image features for the reference trees 
We collected the image data for a reference tree by first estima 
ting a crown envelope using LiDAR data. The crown was 
systematically sampled in 121 surface points, which were pro 
jected to the images. Parallel to this, each point was determined 
if it was visible to the camera or occluded by the tree itself or by 
an adjacent tree (Fig. 3). In addition, an illumination class was 
determined for each point using the LiDAR data in the vicinity. 
Accurate crown envelopes were a prerequisite. The model for 
crown radius was 
r = a 2 + b • h » (1) 
where r is the crown radius at the relative distance x e (0, 0.4) 
down from the treetop. The initial values of the unknowns were 
set using field measurements of tree dimensions and weighted 
least squares adjustment with additional observation equations 
for a and c (to constrain their values) was used for the solution. 
The mean RMSE of r was 0.35 m in 15627 trees. All envelopes 
were convex with c € (0.01, 0.93). 
Figure 3. Determination of camera-visibility and illumination 
class for the 121 crown points. Each tree had 10 layers of 
points, and the 12 points in each had a 30° azimuth offset 
between points. The first point was always aligned in the 
direction of the solar azimuth. Two rays were cast - one to 
wards the camera and another towards the Sun. LiDAR points 
were treated as 0.7-m-wide spheres (grey circles) and tested for 
intersection. The vector angles between the crown surface 
normal and the two rays defined the self-occlusion and self-sha 
ding. The example shows a camera-visible, neighbor-shaded 
point. 
The acquisition of pixel data was repeated for the 15627 trees 
in 54 MS images representing different strips (15), radiometric 
corrections (2 or 3 per strip), and view configurations (1 or 2). 
Cloud screening was done in XY polygons of clouds and sha 
dows. Crowns were sampled in 121 crown surface points (Fig. 
3) that were at different relative heights symmetrically around