3.3 Discriminant analysis 
A linear discriminant analysis will discriminate 
81 % of the thinned stands, table 4. The analysis 
is made with cross validation and with equal prior 
probabilities. TM 1 did not contribute to the 
discrimination and was therefore excluded. TM 7 is 
by far the most important channel. It will alone 
classify the population with a total proportion 
correct of 0.88, compared to 0.91 when five 
channels are used. 
The discriminant analysis gives the following 
optimal linear combination of differences from the 
TM channels, for discrimination between thinned 
and unthinned stands: 
L = 2.40TM7+0.63TM5-0.16TM4+0.61TM3-1.61TM2 
If L > -4.04, classify as thinned, else unthinned. 
Table 4. Confusion matrix from discriminant 
analysis between observations before and after 
thinning, based on changes in TM channels 
2,3, 4,5, 7. 
Classified as : 
True group : 
not thinned 
thinned 
not thinned 
100 
7 
thinned 
5 
29 
Prop, correct 
0.95 
0.81 
Total proportion correct classified : 0.91 
Of the 12 misclassified observations, 6 are due 
to that the thinnings have been classified 1 year 
earlier in 3 stands. All the 5 misclassifications 
in the un-thinned group belong to the first year 
before the reported thinning. Thus, most of the 
errors are probably due to that the cutting 
started before the month reported. 
4. DISCUSSION 
The findings by Hame (1986), that thinning 
cuttings on an average give increased response in 
TM 3 and decreased in TM 4 are confirmed. 
Moreover, an average increase of around 0.5 DN 
was found in TM 1 and TM 2. Most important is an 
increase of 4 DN in TM 5 and 2 DN in TM 7. 
When more than 40 % of the forest is cut, the 
radiance will increase quickly for most channels. 
This is probably due to reduction of internal 
shadows in the stand. 
For the interval 20 % - 40 % thinning grade it is 
not possible to explain the measured DN 
differences as a function of differences in 
thinning grade. One explanation might be that 
factors causing increased and decreased radiance 
in this interval balance each other. 
The tree species composition of the stands has a 
strong influence in the interval 20 - 50 % 
thinning grade. Thinnings in pine forest give a 
much higher DN response. The Pine has contrary to 
Spruce and Birch, generally only the canopy on 
the upper half of the stem. The response in the 
near infrared light is naturally also affected by 
the change in proportions of deciduous forest. 
Examples of other factors that might cause 
spectral differences between the thinnings are the 
amount of logging roads, the field and bush 
vegetation, the sun angle at the acquisition 
after the thinning and the influence of mixels at 
stand borders. 
The possibility to separate thinned stands from 
not thinned stands depends on how much the 
variance of the residuals in the normally 
developed stands can be reduced. This is a 
function of: 
- calibration methods, 
- geometric errors, 
- stand delineation procedures, 
- season and time span between the registrations, 
- technical quality of the satellite data, 
- stand size. 
Of these factors, the first three might be further 
studied in this project. The calibration methods 
applied here were straight forward. The results 
might be tuned with e.g. the use of different 
calibration functions for different types of 
forest, and estimation of calibration functions 
with multiple regression. In a operational system 
must the geometric registration be automatic, and 
carried out with sub-pixel accuracy, using 
correlation techniques. 
No attempt has been made to find automatically 
and delineate the thinned stands. This might be 
possible if the data are carefully treated 
through all processing steps and standwise 
segmentation techniques are used (Hagner 1990). 
Another application of the results is in a 
standwise monitoring system. The DN mean values 
for previously mapped forestry compartments can be 
yearly compared. In such a system digitally stored 
map and register information will also contribute 
in the calibration procedure (Olsson 1990). 
To conclude, there is a possibility to develop 
procedures for the detection, and maybe even for 
mapping of thinnings with Landsat TM data. But the 
data must be optimally treated for any chance of 
success. For normal thinning grades, it will 
probably not be possible to estimate the grade of 
thinning. The changes caused by thinnings will 
also be difficult to separate from changes caused 
by various forest damages, as for example storm, 
fungi or snow damages (Olsson 1989,1990). 
ACKNOWLEDGEMENTS 
This research was financed by the National Board 
of Forestry. The Swedish Space Corporation 
contributed with six geometrically precision 
corrected Landsat TM scenes. The National Forest 
Inventory contributed with the majority of the 
field work, and with functions for calculation of 
the forestry parameters. The work has been 
inspired by discussions with, among others, 
Professors, R. Sylwander, L. Bondesson, Dr. R.W. 
Thomas and Mr 0. Hagner. The forestry company 
MoDo, The Swedish Forest Service, and the County 
forestry boards around Umeâ contributed with 
information about thinning cuttings made by them.