345 
In; Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
the crown envelope. We did not strive for complete pixel lists or 
image patches per tree, but samples of them. 
A total of 158 features were derived for each tree, using the pi 
xel data of the camera-visible crown points. Crown surface 
points belonged to an illumination class: Sun-lit (SL), self-sha 
ded (SS), neighbour-shaded (NS), and neighbour-and-self-sha- 
ded (BS). Because of the viewing geometry and the peaked sha 
pe of the crowns, several points often mapped to the same pixel. 
Duplicates were filtered. The features for each tree and illu 
mination class and the RED, GRN, BLU, NIR, and NDVI bands 
were: min, max, mean, sdev, and quartiles ql-q3. The topmost 
and lowest SL pixels were stored as separate features and band 
ratios were also computed. 
2.7 Variables describing the view-illumination geometry 
To describe the view-illumination geometry we used the phase- 
angle, azimdiff, and offNadir angles. Phaseangle, [0°, 180°] is 
the vector-angle of the camera and Sun vectors. Azimdiff, [0°, 
180°] is the azimuth difference of the camera and Sun vectors 
and it is 0° for perfectly front-lit trees and 180° for back-lit 
trees. OffNadir was the angle between the plumb line and the 
camera vector. Fig. 4 illustrates the sampling of the view-illumi 
nation geometry. 
offNadir 
Figure 1. Distribution of azimdiff x offNadir observations (N = 
202136) for all trees in all 19 strips/views. Division 
between front-, side-, and back-lit trees in azimdiff is 
shown by the vertical lines. 
2.8 Statistical tools and classification methods 
We used analyses of variance and covariance (ANOVA, 
ANCOVA). Classification trials were done with the quadratic 
discriminant analysis (QDA) with equal prior probabilities. 
Overall classification accuracy and the simple Kappa were the 
performance measures. 
3. EXPERIMENTS 
3.1 Evaluation of radiometric corrections and reflectance 
anisotopy in trees 
Atmospheric effects, the changing solar elevation and the 
reflectance anisotropy of trees influence the pixel values in ASR 
images. Ideally, only the variation due to the reflectance aniso 
tropy remains in the ATM images. In the FULL data, which 
combines a BRDF-correction with the ATM correction, the ref 
lectances should be corrected also for a general anisotropy. 
We first examined the ATM and FULL corrections for the 
coefficient of variation (CV = sdev / mean) of the intraspecies 
reflectance. The effects of the ATM and FULL corrections in 
CV were strongest in the BLU band, where the relative CVs 
ranged from -50% to +59% (compared to ASR data). In all 
bands, the effects were strongest in strips that were flown per 
pendicular to Sun. The FULL correction did not completely 
correct for the anisotropy (Fig. 5) although it produced images 
that are well-suited for seamless mosaicking. 
SL mean RED 
Figure 5. Scatterplot of averaged Sun-lit image data from the 
RED (strip 0833/N00A) band and the offNadir angle. 
FULL data. Strip 0833 was flown almost perpendicular 
to the Sun, where the BRDF effects are strong. Red = 
pine, green = spruce, blue = birch. 
We also compared 1-4 km strips from different flying altitu 
des. The mean SL ATM reflectances per species varied up to 
38% and the differences were explained by the changes in the 
view-illumination geometry between strips. In diffuse light, the 
relative differences were smaller. In well-defined targets, for the 
same strips, the differences were less than 10%. In two over 
lapping 1 km strips having a 22-minute temporal mismatch, the 
mean reflectances by tree species varied 2-15% depending on 
the band and species. The well-defined targets in these strips 
showed reflectance differences of below 2%. In an analysis that 
combined 15 stripxview combinations, the well-defined targets 
showed standard deviations of less than 8% for the relative dif 
ferences. When restricting to reflectance tarps only, the sdevs 
were less than 6%. The results show that for well-defined tar 
gets, the precision in the ATM data was high, while the diffe 
rences observed in trees were mostly due to strong reflectance 
anisotropy and to the naturally higher variation. 
SL mean RED SS mean RED 
Figure 6. Distribution of RED SL (left) and SS (right) ATM 
reflectances (0-0.1) and the phaseangle (28°-88°). All 
19 strips/views. Pine (top, N=80073), spruce (middle, 
N=93415), and birch (bottom, N=28648). 
i 
i