397
2.6 Regression analysis
Regression functions for estimation of stand
characteristics from spectral signatures, were
constructed by means of regression analysis of the
NFI-data, described in section 2.3, and corre
sponding Landsat TM spectral signatures. The
statistical software package Minitab (Version
6.1.1 for VAX/VMS) was used for the analysis.
Regression functions were constructed for the
following stand characteristics: Volume/hectare,
mean diameter, mean age, and tree species mixture
(relative proportions of species). The selection
of predictor variables to be included in the
models was based on residual studies and the
adjusted coefficient of determination (R-sq(adj)).
The regression models used were of the type shown
in equation 3.
Log(y)=b 0 +b 1 *x 1 +b2*x 2 ...+b n *x n (3)
where: y = volume/hectare, age, diameter, etc.
x^ = band intensities, band products, band
ratios, etc.
The accuracies of t-ratio segmentation and manual
delineation were found to be very similar for most
stand variables except for mean diameter
(Figure 3). However, the average region size was
10% smaller for the segmentation result. The CFU-
time (VAX 8530) needed for the t-ratio segmen
tation of 1100 hectares was less than two minutes.
The quality of the results obtained by computer
segmentation was also confirmed by field inspec
tion.
STAND VARIABLES
Correction factors for logarithmic bias were
calculated for all regression functions as the
ratio of: mean of real observed values divided by
the mean of real predicted values (Equation 4).
Mean of observed values
Correction factor (4)
Mean of predicted values
Standwise estimates of characteristics, for any
stand within the satellite scene, may be calcu
lated by applying the regression functions to
individual pixel signatures within the stand and
multiplying the averages obtained with the correc
tion factors for logarithmic bias.
3 RESULTS
3.1 Stand delineation
3.1.1 Segmentation. The result of a computer
segmentation using the t-ratio algorithm described
in section 2.5, was compared to the manual deli
neation of test site no. 1, obtained by visual
interpretation of aerial photos and field check
ing.
Three bands of composite SPOT XS,PAN data were
used as feature bands. The overlay defining diff
erent land use and administrative regions was
digitized from public maps at scale of 1:10 000.
Both individual pixels and results of low level
segmentation (Directed trees algorithm) were
tested as definitions of initial regions. If the
average size of input regions was fairly small
(i.e. less than 10-15 pixels), more or less equi
valent results were obtained with the two app
roaches. The appropriate level for the t-ratio
threshold was simply determined on a trial and
error basis.
The accuracy was quantified as the preportion of
stand characteristic variance explained by the
delineation (R-sq, ANOVA) The statistics were
calculated from the sample plot grid of test site
no. 1. Only plots with forest older than 15 years
were used in the analysis.
Figure 3. Relative effectiveness of stand deline
ation (ANOVA, R-sq %), for t-ratio segmentation of
composite SPOT XS,PAN data compared to visual
interpretation of black and white aerial photos at
scale 1:30 000 and field checking.
3.1.2 Interactive editing. A raster based
system for interactive editing on an image back
ground, has been developed at the Remote Sensing
laboratory in Umea. The software is written in
Pascal and presently runs on a VAX-hosted image
processing system (Gould IP8500). The editing
system has a very user-friendly menu interface
(Eliasson, 1989).
3.2 Estimation of stand characteristics
Estimates of stand characteristics were calculated
for each of the 80 reference stands described in
section 2.2.2. The estimates were obtained by
applying regression functions constructed from NFI
and landsat data to all pixel intensities within a
stand, except for the edge pixels. Stand edge
pixels were excluded in order to avoid mixed
signatures and to reduce the effects of geometric
distortions due to topography.
The precision of estimates obtained from satellite
and NFI-data were found to be comparable to these
obtained by subjective field inventory, except for
tree species mixture. A substantial reduction of
estimate error was found when combined estimates
were calculated, using both satel 1 ite-NFI plot
regression functions and subjective field inven
tory estimates. (Figure 4).