In: Wagner W., Szekely, B. (eds.): ISPRS ТС VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
i 
ally positioned the treetops using photogramme trie techniques 
(instead of automatic tree detection) and used field measure 
ments (instead of LiDAR measurements) to define the initial 
approximation of the crown envelope to be fitted in the LiDAR 
data. The RMSE of 0.35 m for radius suggested that the crown 
models were rather accurate. However, the XY position of the 
crown model certainly had small offsets due to measurement 
errors. Since real crowns are seldom rotationally symmetric, it is 
evident that the models were imprecise. Thus, when the crown 
model was sampled in surface points, some points mapped to a 
pixel belonging to a neighbouring tree or background. These 
should result in random noise. We cannot, however, exclude 
possible systematic errors. The ray-tracing based determination 
of shading and occlusion by neighbours, where LiDAR points 
were interpreted as opaque spheres, worked satisfactorily based 
on the analysis of the different illumination classes. Visual 
examination confirmed that the class was mostly correctly 
resolved. The division of canopy points between direct and 
diffuse light is an ill-posed task, since real crowns are semi 
transparent, fractal objects. 
4.4 Aim III: Evaluation of the ADS40 line sensor data for 
tree species classification 
The ATM and FULL corrections worked well within the limits 
of the underlying models. The atmospheric correction was vali 
dated with well-defined targets that show only small BRDF 
effects. The results suggest that the precision of the ATM data 
was better than 10%, with NIR being the most precise and BLU 
the least precise. The analysis with the ATM data, where nadir 
views from different flying altitudes were compared, showed 
that the mean reflectances by species could vary -7%-+38%, 
depending on the band. The results showed that trees are sensi 
tive to changes in the view-illumination geometry, which resul 
ted in effects larger than the reflectance differences between 
species. The sp. classification trials demonstrated that the ATM 
and FULL corrections did not improve the classification perfor 
mance. One reason is that the standard atmospheric correction 
theory does not treat shadow pixels correctly. In ANCOVA, 
62-79% of the total SL reflectance variation in the BLU band 
of the ATM data that combined all strips and views was 
explained by a model that had azimdiff, phaseangle, azim- 
diffxphaseangle, and the strip/view class variables. In NIR, only 
15-18% of the variation was explained by the same model, and 
the R 2 were higher in pine and spruce compared to birch. The 
results suggest that the anisotropy varies between species, and is 
strongest in the visible bands. A single BRDF-normalization for 
pine, spruce, and birch will likely fail. The BLU band features 
were strong predictors of the species. The proximity effects de 
tected here differ from classical atmospheric scattering induced 
adjacency effects. To the best of our knowledge, this was the 
first study to show these effects, which affected the mean 
reflectance 1-17% in VIS bands and up to 33% in NIR. Effects 
by tree age and siteindex explained less than 5% of the 
reflectance variation with the strongest influence in NIR. Plot 
effect explained 1-19% of the reflectance variation and the ef 
fects were again strongest in NIR. The results of intracrown ref 
lectance variation showed that the trends in the mean reflectan 
ces of crowns can be traced to intracrown variation, which 
could be used in improving the feature extraction. The 3 and 4 
km nadir data showed best-case classification accuracies of 
80%, which are higher than those simulated in Heikkinen et al. 
(2010). The fact that the reference trees were scattered across a 
large area and classified as a whole (vs. stand or image frame), 
resembling to a real forest inventory setup, is very important 
practical aspect, and illustrates that ADS40 in general could be 
very cost-efficient for tree species recognition to complement 
LiDAR data. 
4.5 Suggestions for future research 
The results in intracrown variation suggest that a different 
sampling strategy or weighting of the crown pixels might 
improve the features. The anisotropy was similar in pine and 
spruce, suggesting that an anisotropy correction would norma 
lize the reflectances. Otherwise, in noisy data with high 
intraclass variation and small interspecies differences, the 
enhanced species recognition algorithm should measure the 
anisotropy in multiple views and use it to predict the species. 
This could be possible in frame images having high overlaps. It 
will be interesting to compare the performance of the ADS40 
with other cameras in sp. classification and to complement the 
image features with LiDAR data to evaluate the combined 
performance. Also, we did not combine the two CCD views in 
ADS40. Heikkinen et al. (2010) suggested increasing the 
number of bands to five and reducing their spectral width in 
ADS40. 
5. REFERENCES 
Beisl, U., Telaar, J., Schonermark, M. V., 2008. Atmospheric 
correction, reflectance calibration and BRDF correction for 
ADS40 image data. IAPRS, Beijing, Vol. 37, part B7, pp. 7-12. 
Deering, D.W., Eck, T.F., Banerjee, B., 1999. Characterization 
of the reflectance anisotropy of three boreal forest canopies in 
spring-summer. Rem Sens Env 67 (2), pp. 205-229. 
Fraser, R. S., Ferrare, R. A., Kaufman, Y. J., Markham, B. L., 
Mattoo, S., 1992. Algorithm for atmospheric corrections of 
aircraft and satellite imagery. Int J Rem Sens 13 (3), pp. 541— 
557. 
Holmgren, J., Persson, A., Soderman, U., 2008. Species identi 
fication of individual trees by combining high resolution 
LiDAR data with multi-spectral images. Int J Rem Sens 29 (5), 
pp. 1537-1552. 
Heikkinen, V., Tokola, T., Parkkinen, J., Korpela, I., Jaaskelai- 
nen, T., 2010. Simulated Multispectral Imagery for Tree Species 
Classification Using Support Vector Machines. IEEE Trans 
actions Geos Rem Sens 48 (3), pp. 1355-1364. 
Korpela, I., 2004. Individual tree measurements by means of 
digital aerial photogrammetry. Silva Fennica Monographs 3, pp. 
1-93. 
Larsen M., 2007. Single tree species classification with a hypo 
thetical multi-spectral satellite. Rem Sens Env 110 (4), pp. 523- 
532. 
Li, X., Strahler, A. H., 1986. Geometric-optical bi-directional 
reflectance modeling of a coniferous forest canopy. IEEE Trans 
actions Geos Rem Sens 24, pp. 281-293. 
Markelin,L., Honkavaara, E., Beisl, U., Korpela, I., 2010. Vali 
dation of the radiometric processing chain of the Leica airborne 
photogrammetric sensor. ISPRS Vienna 2010. 
Martonchik, J.V., Bruegge, C.J. , Strahler, A.H., 2000. A review 
of reflectance nomenclature used in remote sensing. Remote 
Sensing Review 19, pp. 9-20. 
347