Because of regression calibration the intensities in both
AX, = studentized difference of Xe {AM, ASD}
dependent and independent images were scaled to the level of AXı = difference of X e{AM,, ASd, } expressed in dn on
independent image. When more generic training data was channel i
considered, the regression calibrated difference observations RMSE = root mean square error of the calibration model
were studentized (in the sense of Weisberg 1985) according to channel i
the leverage and regression error (equation 7) (e.g. Olsson lev, leverage of stand x mean intensity on channel i
1994). (Weisberg 1995).
AX rure (7)
RMSE, \[1+ lev,
3. RESULTS
The coefficients of determination in calibration models are
presented in table 3. Studentized
4p difference
Table 3. The coefficients of determination of the models -
applied in calibration |
Coefficient of determination at different sol
calibration intervals I
™ H93-90 H92-90 H93-92 .. ppm:
channel xL |. . |
1 0.66 0.79 0.84 TE di
2 0.66 0.79 0.80 |
3 0.59 0.82 0.75 10|-
4 0.93 0.94 0.93 | mu
5 0.92 0.93 0.94 | I“ Standard
6 0.83 0.90 0.73 0 deviation
7 0.80 0.85 0.88 Br ne 9
The RMSE's of the regression calibration varied from 0.22 on
channel 7 to 0.88 on channel 6 expressed in eight bit digital
number (DN) scale.
After regression calibration and studentization, the timing of
the changes within the three-year interval did not affect to the
separability of the change classes (Figures 1, 2, 3 and 4). The
only exception was the regeneration cut where the earlier
treatments seemed to have larger spectral responses compared
to later ones. However, this was probably due to the small
number of observations.. The class 'untreated' did not differ
from zero according the T-test at 1% risk. The decreasing
response in 'clear cut' is due to soil preparation which was
combined with two earlier intervals, but had not yet been
accomplished in the later interval (Figures 1 and 2)
The separability of different forest management treatments
based on their spectral response was compared between
uncalibrated and regression calibrated/studentized difference
images with three-year interval between images. According
to T-test all the treatment classes were separable at least on
one channel based, on differences of stand mean intensities
only. The best channels for separating different treatments
were TM 2,3,5 and 7. None of the change classes were
separable on channel four. The calibration error was relatively
small compared to spectral changes due to treatments
(Figures 1,2,3 and 4). However, the channels from which the
different treatments were separable varies between
uncalibrated and regression calibrated/studentized calibration
alternatives.
Implementation period
Figure 1. Spectral change of clear cut on the difference image
H93-90, TM channel 3.
Studentized
150 difference
100} |
50H
M calib
error
| Standard
0 deviation
90-91 91-92 92-93
Implementation period
Figure 2. Spectral change of clear cut on the difference image
H93-90, TM channel 5.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
Figt
ima,
The
supe
spec
diffe
the s
sam
char
eo
20
15
10
Figu