In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
For the analysis of anisotropy, we combined the ATM data in
the 19 strip/views giving over 200 000 observations. Figure 6
shows the anisotropy in ATM data (RED band) as a function of
the phaseangle. The high intraspecies variation (CVs 12-38%
over all band and species) is also obvious from Figure 6 as well
as the small intraspecies differences in the mean reflectance. In
diffuse light, there was no trend in high values of phaseangle
i.e. for the back-lit trees. The anisotropy was stronger in visible
bands compared to NIR (with high reflectance). Also, pine and
spruce seemed to differ from birch. We tested if strongly back
lit trees would show high maximal reflectances due to forward
scattering (of foliage). Such phenomenon was not observed.
3.2 Inter- and intratree reflectance variation
We studied proximity effects i.e. the influence of the immediate
neighborhood to the observed reflectance. The effects were
considerable in the NIR band, where adjacent birch trees caused
an effect up to +33%. In the visible bands the effects were
within ±10% being stronger in the diffuse light conditions.
In ANCOVA, reflectance variation was explained by phase-
angle, azimdiff, phaseanglexazimdiff, the strip/view class (N =
19) variables, and age or siteindex (class). The models were
fitted to the combined ATM data of all 19 strips/views. The
results showed that 1-4% of the variation was contributed by
the age of the trees with strongest effect in NIR and birch. The
siteindex explained 1-3% of the NIR and RED band
reflectances. When the stand was used as explanatory class va
riable, 1-19% of the reflectance variation could be contributed
to the stand effect, with strongest effects in NIR in all species.
Strip/view
BAND
SL
SS
NS
BS
0825/
B16A,
front-lit
trees
BLU
1
0.92
0.87
0.84
GRN
1
0.75
0.67
0.59
RED
1
0.73
0.66
0.57
NIR
1
0.70
0.62
0.51
0818/
B16A,
back-lit
trees
BLU
1
0.96
0.90
0.87
GRN
1
0.89
0.71
0.63
RED
1
0.92
0.72
0.64
NIR
1
0.85
0.63
0.51
Table 1. Relative mean ATM reflectances in illumination
classes (all species combined) in 3 km N-S oriented
strips 0818 and 0825.
Offset
to Sun
Species
BLU
GRN
RED
NIR
0°
pine
1
1
1
1
30°
“
0.99
0.98
0.97
0.97
60°
“
0.97
0.92
0.90
0.91
90°
“
0.94
0.84
0.81
0.83
0°
spruce
1
1
1
1
30°
“
0.98
0.95
0.95
0.96
60°
“
0.95
0.85
0.84
0.84
90°
“
0.91
0.75
0.73
0.73
0°
birch
1
1
1
1
30°
“
0.99
0.96
0.96
0.95
60°
“
0.95
0.87
0.87
0.85
90°
“
0.92
0.77
0.77
0.75
Table 2. Relative mean reflectances of crown points in the SL
illumination class with varying solar azimuth offset.
Strip 0825/B16A and ATM data with front-lit trees.
Values are normalized to the 0° azimuth offset class
Intracrown reflectance variation was examined using the crown
point data per tree. Table 1 shows the mean values of ATM
reflectances in the for illumination classes for front- and back-lit
trees. Table 2 shows the effect due to solar azimuth offset. The
determination of the illumination classes succeeded based on
data in Table 1, where the order of brightness was SL > SS >
NS > BS. The relative difference of SL and SS was not preser
ved for front- and backlit trees. Table 2 shows how the brightest
pixels are found in crown points that are towards the Sun and
how the effect is smallest in BLU.
3.3 Feature selection and tree species classification trials
We confined to a suboptimal manual feature selection, which
combined ANOVA and correlation analysis. Seven features for
the classifications were: SL_sdev_GRN, SL_q3_RED, SS_-
mean_BLU, SL_q3_NIR, SL_q3_NDVI, SL sdev NIR, and
SL mean BLU. The BLU band was a strong predictor, which
calls into question the severity of the possible imprecision in the
atmospheric correction of the BLU band (Section 3.1).
Monoscopic nadir data was used with 3 and 4 km strips. The
classification accuracies by strip were 72-80% (k = 0.56-0.67)
with ASR (73-80%) and ATM (75-80%) producing higher ac
curacies than FULL (72-78%). Accuracy varied between strips
and was lower in the E-W oriented strips (72-76%) compared
to the N-S strips (75-80%). Leave-tree-out and leave-plot-out
validation methods produced similar accuracies as well as
teaching the QDA-model with randomly selected subsets of
3x300 trees per species, which was 5-10% of the number of
trees in the validation data set. Pine was classified most accura
tely (83-85%) in the N-S strips, with spruce at 73-76% and
birch at 76-79%. In a trial, where 182982 observations in the
1-4 km ATM data were combined, the accuracy was only 62%
in leave-one-out cross-validation. By restricting to three 3 km
N00A views, accuracy was 72%. These results show how the
anisotropy hampers classification if the view-illumination geo
metry is not restricted to the use of single N00A strips. When
the N-S oriented strip 0825 was used for teaching and the E-W
oriented strips 0843 and 0852 were used for validation, the
accuracies were 54% and 68%, which demonstrates the findings
of Section 3.1.
4. DISCUSSION
4.1 Confines
The presence of clouds may have affected the results. Also the
timing, August 23, was quite late. The number of trees was very
high. However, trees younger than 25 years were mostly mis
sing as well as dominated trees, which remain unseen. The num
ber of plots (121) overestimated the number of stands, because
of the large size of regeneration areas in Hyytiala.
4.2 Aim I - Sensor model implementation
We needed to implement the sensor model to do this research.
However, this is a technical problem and many modem com
mercial digital photogrammetric workstations have the ADS40
sensor model implemented allowing standard photogrammetric
procedures and workflows.
4.3 Aim II - Crown modelling and determination of illumi
nation classes
We improved an existing method for crown modelling in
LiDAR data by imposing constraints on the parameters, which
enhanced the solvability of the non-linear regression. We manu
