2.2. Digitized map data 
Digital map data will be used for improving the accuracy of estimates and to separate 
forest and non-forest, land from each other. The spectral response of peatlands differs 
from that of the mineral soils with the same growing stock according to our earlier 
studies. Further, some peatlands can not be separated from mineral soils. Therefore, 
digital peatland information will be used in order to improve the accuracy of estimates. 
These digital data are provided by the National Board of Survey. The peatland data, 
digitized from a generalization of a base map with the scale of 1 : 100 000, are now 
available for the whole country. 
Agricultural areas and roads will be digitized from base maps having a scale of 
1 : 50 000 (if the preliminary tests show these data to be useful). A combination 
of a numerical interpretation and digital map information will be used in classifying 
agricultural areas, because the map data are not necessarily up to date. 
Urban areas can be obtained from the house register provided by the statistical 
centre of Finland. The coordinates of each house in Finland are known. A digital urban 
area mask can be produced from this information. 
Water areas could be obtained from base maps but they can be obtained relatively 
reliably also from satellite images. 
Some administrative information such as community boundaries and, in the future, 
boundaries of forest holdings will also be used in t he digital form in order to differenta- 
tiate computation units. 
2.3. Ground truth data 
The recent NFI data can serve as ground truth data. In the future sampling design, 
the needs of satellite image interpretation must be taken into account more thoroughly. 
Such problems as 1) the sampling intensity of permanent and temporary plots, 2) 
the rotation period of inventory, and 3) the size and the form of the sample plot and of 
the cluster will have to be addressed. Choices must be made, for example, between one 
relatively small or several even smaller plots in the same stand. (The satellite image 
information is essentially standwise information rather than sample plot information.) 
3. IMAGE INTERPRETATION 
The image analysis consists of preprocessing of t he image (removal of noise, striping, 
etc.), choice of features, classification, and postprocessing (generalization). All inventory 
variables should be estimated. One possibility is to use some lion-parametric method 
such as nearest neighbor classification (see Appendix I ). This method can also lie 
referred to as a 'nearest neighbor fuzzy classification'. Each interpretation class consists 
of one sample plot (one or several pixels depending on what kind of features will be 
used). Each pixel to be classified will be shared between several classes according 
to Appendix 1. The shares are inversely proportional to the squared feature space 
420