In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
342 
VARIATION AND ANISOTROPY OF REFLECTANCE OF FOREST TREES IN 
RADIOMETRICALLY CALIBRATED AIRBORNE LINE SENSOR IMAGES - IMPLICA 
TIONS TO SPECIES CLASSIFICATION 
I. Korpela *, F. Rohrbach 
Faculty of Agriculture and Forestry, University of Helsinki, POB 27, 00014 UH, Finland - (ilkka.korpela@helsinki.fi) 
Commission VII 
KEY WORDS: Forestry, Radiometry, Modelling, Calibration, Pushbroom, Multispectral, LiDAR, 
ABSTRACT: 
In Scandinavia, the conventional method of measuring trees is giving way to applications, which combine in-situ and airborne opti 
cal data, LiDAR in particular. Tree species (sp.) classification is a crucial sub-task and is solved with an insufficient reliability. The 
continuously varying view-illumination geometry hampers image-based solutions. Line sensors provide a selected subset of the 
possible view-illumination geometries, but have not been tried for the task in Scandinavia. We examined the variation and anisotropy 
of reflectance in trees, using radiometrically calibrated multispectral ADS40 data. An experiment in Finland (61°50’N, 24°20'E) that 
consisted of 121 plots and 15197 pine, spruce, and birch trees was imaged from 1, 2, 3, and 4 km altitudes. Leica XPro was used for 
producing different image data including the at-sensor radiance (ASR) data, atmospherically (ATM) corrected, and a combined 
BRDF- and atmospheric correction (FULL). Tree crowns were modelled in LiDAR data, and the resulting envelopes were sampled 
in the images in 121 points. Using the geometry of the crown envelope and the adjacent LiDAR points to model the geometry of the 
neighbourhood, camera-visibility and illumination class (Sun-lit, self-shaded, neighbour-shaded) was determined for each point. 
Using the pixel data of the crown points, different statistical features were derived for each tree. The radiometrically corrected image 
data did not reduce the intraspecies coefficient of variation, and in sp. classification trials, the ASR data provided equal or better 
results. The precision of the ATM data was evaluated to be better than 10% with the NIR band being most precise and the BLU band 
least precise. However, the BLU band was a strong predictor of tree species. Reflectance anisotropy of pine and spruce differed from 
birch, and it was strongest in the visible bands and varied up to ± 40% in nadir lines flown nearly perpendicular to the Sun. 
Reflectance of crowns in diffuse illumination showed lower anisotropy and features derived in these data were strong predictors of 
species. We observed notable proximity effects in the NIR band, where the species composition of the adjacent trees affected the 
observed reflectance of the target tree up to 33%. Intracrown reflectance variation was examined for crown points oriented towards 
or away from the Sun on different relative heights. Age dependencies were observed in NIR and NDVI, where age explained up to 
5% of the reflectance variation, and the dependency was negative. Site fertility was correlated with NIR and NDVI, and the overall 
stand effect explained 1-19% of the reflectance variation by band and species. This elucidates, why also the tree species classi 
fication accuracy varied considerably between stands. Classification accuracy for pine, spruce, and birch was 72-80% in quadratic 
discriminant analysis, when features of both Sun-lit and diffuse light were used as predictors. Best-case accuracies of 76-80% were 
achieved using 3 and 4 km monoscopic data, which shows the high potential of the ADS40 line sensor. 
1. INTRODUCTION 
In Scandinavia, the conventional method of measuring trees is 
giving way to applications, which combine in-situ and RS data. 
Here, the introduction of airborne LiDAR was a breakthrough. 
The need for aerial images is a topical question amongst Scan 
dinavian foresters to whom species information is crucial on 
technical, economic, and ecological grounds. Separation of 
Scots pine, Norway spruce, and birch is essential for forestry in 
Finland. Classification accuracies of above 90% are considered 
adequate for practice. Recently, low-altitude, high-density 
discrete-return LiDAR data were tested for tree species discri 
mination with accuracies saturating at the 85-90% level. Owing 
to the monostatic view-illumination configuration, LiDAR sig 
nal is largely free from view-angle effects that have been repor 
ted to hamper image-based tree species recognition. In Sweden, 
an accuracy of 84% was reported in 1 km altitude, digital frame 
camera data (60 cm GSD DMC) (Holmgren et al 2008). 
The validation procedures used for estimating species classifi 
cation accuracy should provide realistic demonstrations of the 
performance. The leave-one-out cross-validation is often used. 
It results in optimistic evaluation of the performance, omitting 
the spatial autocorrelation in forests and the similarity of view- 
illumination effects for neighbouring trees. 
A species classification accuracy of 95% would be adequate 
for foresters in Finland, which is very challenging in airborne 
optical data, because the observations from the above result in 
commission and omission errors in tree detection. If a small tree 
is detected in near-nadir LiDAR, the low solar elevation, <53° 
in Finland, prevents from observing but the tallest trees in direct 
light. Species identification in image-based analysis is hampe 
red by the view-illumination geometry (Fig. 1). According to Li 
and Strahler (1986), the geometric nature of the forest canopy is 
the major factor explaining the strong anisotropy of directional 
reflectance. Anisotropy of the reflectance is a fundamental 
property of most objects, and it means that the observed bright 
ness is dependent on the illumination and observation direc 
tions. For point-like illumination and observer, the dependence 
can be described by the BRDF, depending on two illumination 
and two observation angles. Illumination conditions in forest 
canopy range from direct illumination to combinations of diffu 
se and direct illumination down to the total diffuse illumination, 
* Corresponding author