maps were digitized and overlaid on the 
imagery and used as ground truth data. 
Black and white aerial photos, which was 
taken in 1990 and were 1 to .20,000. in 
Scale, were also used as ground truth. 
3. METHOD 
3.1 Successional Spectral Changes 
Twenty one spruce stands, which were the 
darkest in the same age class in D0708 
and D0309, were selected as training 
areas. Average DNs of training areas from 
each imagery were plotted against the 
stand. ags- in TM3, - TM4. -and /TM5.- Then .an 
exponential curve was fit, and 
successional spectral changes were 
evaluated in each image (Figure 1, Table 
De 
Exponential Curve Fitting: 
DN(CHj,k) — a(CHj,k) * b(CHj,k)* 
exp (-c(CHj, k)*Age) (1) 
where 
DN(CHj,k) is estimated DN in channel j of 
image k. This is referred as ‘average’ 
hereafter. RECHT Xk), D(CHj,k} and 
‘C(CHj,k)'’ are regression coefficients for 
channel j of image k. ‘Age is the stand 
age of training area. 
Relationships between the standard 
deviation of the training areas and the 
age were also checked using equation (1) 
for TM3, TM4 and TMS5 (Table |). 
3.2 Classification 
Averages and standard deviations of DNs 
in TM3, .TM4 and TM5.at every stand. ages 
between 1 and 100 years were estimated as 
training >» data using the exponential 
curves (Table 1). The average standard 
deviation of all training areas was used 
for all ages, when a correlation 
coefficient between the standard 
deviation and the stand age was very poor. 
The minimum distance classifier, which is 
defined as following equations (Takagi 
and Shimoda, 1994), was used .to classify 
pixels into age classes. 
Distance Calculation: 
DST (i,CHj,k) = (DN(i,CHj)-AVR(CHj,k)) 
/ SD(CHj, k) (23 
N 
n 
DSTA(i,k)^- JJ" XBST(i,CHj,kE) (3) 
j=1 
where 
DST(i,CHj,k) is the distance of ‚pixel | 
about stand age k in channel j. DN(i,CHj) 
is DN of pixel i in channel j. AVR(CHj,k) 
is the estimated DN for stand age k in 
channel J using the exponential curves 
(Table. 1). SD(CHj, k) is the standard 
deviation at stand age k in channel jj 
defined by the exponential curves (Table 
26 
Lp. DSTA(i,;kyoiss/distance of'pixel i from 
the training data of age k. n is number 
of channels, where it was three, namely 
TM3, TM4 and TM5. 
Bach. pixel" of the four images was 
classified into an age class, which 
showed the minimum DSTA for the pixel. 
The age class is called spectral age 
hereafter. 
3.3 Evaluation 
Then relationships of spectral ages 
between images were evaluated using 62 
evaluation areas, which were different 
from training areas for exponential curve 
fitting, ' using the regression analysis 
(Figure:s2)uG The relationship between 
Spectral age and stand age was also 
checked (Figure 3). The evaluation areas 
included various age classes and dense or 
Sparse spruce stands. Images of spectral 
ages were created, then they were 
visually compared between each other 
referring the black and white aerial 
photos. 
4. RESULTS AND DISCUSSION 
Spruce dis... one .s-of . common. climax tree 
Species in ‘the boreal forest and is 
evergreen species. On the other hand, 
most of pioneer species in the cool 
temperate forests are deciduous, the 
seasonal spectral characteristics of 
these species are different. Above all, 
since snow covers small undergrowth in 
the boreal forest (spruce) in winter, 
satellite data were almost composed of 
radiation from: spruce and snow. Imsthe 
meanwhile, satellite data were composed 
of radiation from spruce and undergrowth 
vegetation in summer, a comparison of 
summer and winter images might show any 
differences caused by undergrowth or any 
other forest components. Thus winter and 
summer images were compared to know 
possibility $ofr-amonitoring forest using 
images taken in different seasons and of 
detecting forest structure. 
The estimated spectral ages agreed well 
between images except D0309 (Figure 2). 
Among 6 cases, the slope and offset 
appeared 1 and 0 in the regression line 
between D853 and D858. This would mean 
that the spectral . age . derived . from 
different images are quite similar. 
On the other: hand, the combination. of 
D933 and D937 showed a smaller 
correlation coefficient than that of D853 
and D858, the spectral age didn't agree 
well. Such difference was probably caused 
as: follows. There would: be no snow 
attached with spruce crown in D853 and 
D933. However, the smaller DN of D933 
suggested wetter snow surface than D853. 
The intensity may become closer between 
spruce and snow than dry snow condition. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996 
a) 
18C 
16C 
14C 
12€ 
10C 
8C 
6C 
4( 
2( 
Digital Number 
Digital Number 
D: 
March 
August 
March 
July € 
No of 
classi 
deviat