be made. An alternative approach is to use prin 
cipal components, instead of original bands. 
RMSE % 
40 -I 
VOLUME DIAMETER AGE SPECIES 
STAND VARIABLES 
SATELLITE 
FIELD 
COMBINED 
(3) Pixelwise estimates of stand characteristics 
and site type parameters derived from digital 
elevation models and soil maps can be used as 
feature bands, instead of the original satellite 
data. 
4.3 Estimation of stand characteristics 
The results of this study confirms the finding of 
Tcnppo (1986) that the precision of standwise 
estimates for the main stand characteristics, 
obtained from NFI and Landsat data, are of the 
same order as for traditional field inventory 
methods. 
Figure 4. Accuracy for estimation of stand vari 
ables (Relative root mean square error, FMSE%) by 
means of: Landsat and NFI-plot data, subjective 
field inventory, and combined estimation. 80 
reference stands used. 
4 DISCUSSION 
4.1 The integrated inventory design 
Dae to the fact that every stand is visited in 
the field and that manv sources of information are 
used, the outlined method is robust in the respect 
that gross errors can be detected and corrected. 
All types of stand characteristics, needed for 
management planning, can be described with the 
method and not only those detectable by remote 
sensing. 
The availability of NFI-plots within a specific 
satellite scene is sufficient even for use of 
SPOT satellite data (60 x 60 km). The reasons are 
that also the temporary plots (60%) are now being 
geocoded with the same precision as the permanent 
plots and that the NFI sampling density is higher 
in most parts of Sweden. The training procedure 
(i.e. regression analysis of geocoded NFI- and 
spectral data) is simple and objective, and can be 
done before the start of the field inventory 
season by personnel with statistical training. No 
specific efforts are required for atmospheric 
correction since the spectral intensities are 
referenced directly to objective measurements of 
stand characteristics. The derived functions may 
then be used throughout the entire satellite 
scene by several field crews. 
4.2 Stand delineation 
The results show that stand delineation of high 
quality can be achieved by using the t-ratio 
segmentation method and SPOT data. The performance 
is comparable to visual interpretation of aerial 
photos and field checking. Although the results of 
t-ratio segmentation seem promising, the method 
may be improved in several ways: 
The results also clearly show the potential of 
combining remote sensing and subjective field 
inventory estimates. A substantial reduction of 
the error was found for all stand characteristics. 
The same technique can also be applied to include 
other sources of information in the analysis, such 
as data from previous inventories, estimates 
obtained by interpretation of aerial photos, etc. 
4.4 Future development and implementation of the 
method 
Standard personal computers and workstations new 
provide the computing and graphic capabilities 
needed for the implementation of the complete 
integrated inventory method. The implementation 
can to a large extent be based on the use of 
commercial standard software for statistical 
analysis, GIS-routines, image analysis, etc. It is 
important that the system is implemented in such a 
way that it can be used by the field inventory 
personnel on a daily basis. In future development 
of the methodology the main areas of activity will 
be: (1) improvement and evaluation of the segmen 
tation technique, and (2) development and testing 
of routines for field inventory. 
5 ACKNOWLEDGEMENTS 
This study is part of the project "Computer aided 
forest mapping", financed by the Swedish National 
Space Board (DFR) and the Board of Forestry and 
Agricultural Research (SJFR). Satellite data were 
provided by the Swedish Space Corporation. The 
"IRR" editing system has been developed by Mr. 
Anders Eliasson. The field data sets for sites 2-5 
was provided by Mr. Goran Stahl. The field data in 
test site no 1 was surveyed by Mr. Lars Bjork and 
Roger Johansson. Field data collection has been 
supported by the County Board of Forestry. Ms. Asa 
Lindblcm and Mr. Levi Johansson have helped with 
the data processing. Rewarding discussions have 
been held with Prof. Sipi Jaakkola, Prof. Lennart 
Bondesson, Prof. Robert Sylvander, Dr. Randall W. 
Thomas, Mr. Mats Nilsson, Mr. Hakan Olsson, and Mr 
Peter Holmgren. 
(1) The difference in spectral variation between 
regions can be a strong indicator of different 
forest types and should therefore be taken into 
account, by including seme test for difference in 
variation. 
(2) The full multivariate T 2 statistic can used 
instead of the simplified statistic in equation 2, 
which assumes the feature bands to be uncorrelated 
(not true for satellite data). Implementation of 
the multivariate T 2 is rather straight-forward, 
although a sacrifice of computing performance must 
6 REFERENCES 
Eliasson, A., 1989. Integration of satellite 
remote sensing in a forestry oriented geo 
information system. Remote Sensing Lab 
oratory, Swedish University of Agricultural 
Sciences, Sweden. 
Hagner, O. 1989., Computer aided forest mapping 
and estimation of stand character'sties using 
satellite remote sensing. Reroute Sensing 
Laboratory, Swedish University ot Agricul 
tural Sciences, Sweden. 
398