dwise
'S had
(9)
A
(10)
(11)
(12)
ems N
(13)
itation
height
of the
stands
height
everal
tion is
forest
. Since
ry was
e these
(14)
rentory
Juha Hyyppä
The RMSE was divided into two parts, systematic error x and standard error of the estimate s.
Coefficient of determination, R?, was obtained by dividing the sum of squared standard error explained by the method
by the sum of squared errors explained by the average (SSEA)
SSEA - s?
n° 54 (n
SSEA (15)
4 RESULTS AND DISCUSSION
Table 2 summarizes the results obtained for 41 stands. The estimated accuracy of field inventory is better than reported
in Hyyppä et al.(2000a) since the stand size of the applied material is higher than that used by Hyyppä et al. (2000a).
‘ The effect of stand size was corrected according to results reported by Hyyppä and Hyyppä (2000).
Table 2. Summary of the accuracy estimation.
Data Source/Error Mean height Basal Area Volume
Field inventory/Standard error 1.7m 3.0 m/ha 35.8 m’/ha
Field Inventory/Systematic error +0.57 m 0.0 m?/ha +19.3 m’/ha
Laser scanner/Standard error 23m 1.9 m*/ha 16.5 m°/ha
Laser scanner/Standard error-% 13.6 % 9.6 96 9.5%
Laser scanner/Systematic error +2.5m - 9.7 m?/ha - 65 m’/ha
The mean tree height was obtained with 2.3 m standard error (conventional field inventory 1.7 m). The overestimation
can be explained that laser scanner is capable to detect only the trees that can be seen above the and also due to growth
of 2 years. Since the height of each tree in dominant storey can be assessed with 1 m accuracy, the standard error 2.3 m
is most likely due to errors in field inventory that were not taking properly into account when calculating the corrected
mean squared error and errors due to improper use of the segmentation algorithm (very dense forests). Examples of both
are results obtained for sapling and young stands. Even though laser overestimated these stands, there was one stand
differing more than 9 m from average behaviour (field inventory giving too low estimate). Since it has been found in this
study and in previous study (Hyyppi et al., 1999) that laser does not miss trees in dominant layer using as high pulse rate
and low flying altitude as in this study, it is most likely that there has been severe field inventory error in this stand. This
conclusion was confirmed by new field visit. Additionally, it seems that the use of same parameters for all stands in the
segmentation procedure was not justified. Since the tree height values given as input for the segmentation algorithm are
absolute and not relative tree heights, either the method should be revised or the different stand types should be assessed
with different parameters. The segmentation procedure is originally developed for aerial photos and adapted afterwards
for laser scanner data. Therefore, the segmentation procedure may not fully exploit the capability of 3-dimensional tree
height models.
The high coefficient of determination (R’=0.89) obtained for the basal area suggests a rather good capability to find
individual tree crowns by the segmentation procedure, Table 2. The obtained accuracy of 1.9 m/ha (9.6 96) suggests a
better performance than by using conventional forest inventory. However, a large systematic underestimation is due to
improper calibration (Equation (4), segmentation parameters) and due to the fact the only the trees in the dominant layer
were found. The regression-based model converting crown diameter to stem diameter was formed by using only 25
individual tree measurements. The use of tree height in Equation (4) improved the coefficient of determination from
0.72 to 0.89. The results are especially promising since the basal area is the most difficult parameter to assess using laser
scanner.
The estimates for the stem volume summarize the above-discussed results, since the parameters affecting the stem
volume are the basal area and mean height. The results, however, suggest a promising capability for operational forest
inventories giving more accurate estimates (R^—0.98, standard error 16.5 m’/ha, 9.5 %) than using conventional forest
inventory. The correction of the overestimation in the mean height and underestimation in the basal area measurements
resulted in an underestimation of stem volume. That underestimation after correction of parameters in Equation (4)
should be corrected by introducing the diameter distributions of typical forests within the target area. Smaller trees not
visible should be corrected by adding corresponding tree information from these distributions.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 427