Coordinates for the plots were digitized from 
orthophoto maps at scale 1:10 000. The actual plot 
locations were marked during field inventory. 
Only permanent plots were used in this study 
because only the nominal coordinates were 
available for the temporary plots. The figures of 
volume/hectare, diameter, age, basal area etc. 
were updated to the level of 1989, by applying 
growth models developed by Soderberg (1986). 
2.4 Image and map data 
Ihe digital satellite data used are summarized in 
Table 1. Ihe satellite scenes were precision- 
corrected by the Swedish Space Corporation in 
Kiruna using ground control points and orbit 
modeling techniques. Ihe SPOT XS and PAN data used 
at test site 1 were merged into a multispectral 
composite, with spatial resolution of 10 m, using 
the method described by Jaakkola & Hagner (1988). 
The Landsat quarter-scene acquired 1989, was used 
to provide multispectral signatures corresponding 
to NFI-plots. The intensities were extracted 
without correcting for displacements due to eleva 
tion. 
Landuse information and administrative boundaries 
at the test sites were digitized from public maps 
at scale of 1:10 000. Ihe digitized vector data 
were converted into raster overlays for use in the 
segmentation procedure. 
Table 1. Digital image data used in the study. 
Type 
Resolution 
Date of 
acquisition 
SPOT 1 HKV PAN 
10 m 
18 June 1986 
SPOT 1 HKV XS 
20 m 
18 June 1986 
Landsat 5 TM 
30 m 
21 June 1989 
2.5 t-ratio segmentation 
In the case of multiple bands, the statistic ncg*j 
here is calculated as the square—root of summed, 
squared t-ratios, computed for each feature band 
(Equation 2), which is equivalent to Hotellings 
^ the case of uncorrelated bands 
(Manly, 1986). Note that this assumes independence 
among features. 
b 
T 2 = Sum t 2 A 
(2) 
Where: 
t = t-ratio, band j. 
b = Number of feature bands. 
Ihe t-ratio segmentation procedure is started by 
entering input parameters and defining the initial 
set of regions. Ihe input parameters are: minimum 
and maximum region size allowed, a final t-ratio 
threshold, and the number of iteration steps. Any 
digital image data may be used as the source of 
feature bands. Ihe segmentation is guided by an 
overlay band, which can be used to define diff 
erent administrative regions and land use classes. 
Individual pixels, or results of a lew level 
segmentation (e.g. directed trees segmentation), 
may be used as initial regions. In the special 
case of single pixel "regions", when the t-ratio 
is not defined, the pixel is merged with the 
adjacent region (or pixel) closest in feature 
space. 
In order to control the order of merging, so that 
the most similar regions are merged first, the 
merging is done in several iteration steps. A 
temporary t-ratio threshold is initialized at a 
lew starting level and raised for each step, until 
the final threshold level is reached. Note that a 
region can be claimed by, and merged with more 
than one other region during an iteration step. If 
the number of steps is sufficiently large (10-15), 
most regions will be merged properly at the end. 
The t-ratio segmentation method (Algorithm 1) is 
a type of region growing algorithm. Ihe basic idea 
behind the method is that spatially adjacent 
regions should be merged if they can not be separ 
ated with a given certainty. 
A criterion for merging of regions should describe 
the chance for two regions to be of the same type, 
i.e., test the hypothesis that the spectral inten 
sities of the two regions are in fact observations 
on the same population. Hence, the absolute dis 
tance in feature space between regions must be set 
in relation to the population variance and number 
of observations (pixels). A criterion with the 
properties described above is the well-known 
t-ratio (Manly, 1986) (Equation 1). 
For each step, the following operations are per 
formed in parallel for all regions: A temporary 
threshold is calculated. Each region is compared 
to all adjacent regions and the one closest in 
feature space is selected. If the t-ratio for the 
crurrent and selected region is less than the 
temporary threshold, and the total size of the two 
regions does not exceed the maximum size allowed, 
the two regions are listed for merging. At the end 
of each pass, all regions listed are merged and 
statistics are calculated for the new regions. 
When no more regions are merged, the next itera 
tion step is started. Ihe iteration ends when the 
final t-ratio threshold is reached. Finally, all 
remaining regions smaller than the minimum size 
are merged with the adjacent region closest in 
feature space. An example of t-ratio region seg 
mentation is shewn in Figure 2. 
t-ratio 
x i - x 2 
( s 2 i + s2 x )0-5 
n l n 2 
(1) 
where: X^ = Mean of region i 
s 2 i = Variance region i 
ni =number of pixels region i 
395