Dees, Matthias 
  
Out of the 817 sample plots 735 that belonged to the state forest have been used. The results of the estimates for the 
main attributes is given in table 3. The estimates are given for the main attributes, the total volume, the area and volume 
parameters of the main broad-leaf tree species group and the main coniferous tree species group. All main attributes are 
estimated with smaller sampling error. The reduction of error varies from attribute to attribute. If the total volume is 
given the highest priority, the potential to reduce the sample size due to estimating with stratification is 26%, or roughly 
25%. By reducing the grid density in one direction by 0.5 , this reduction can be easily achieved. Further analysis 
planned in the study comprises an alternative definition of strata (4 age classes mixed with 2 species type classes 
[broad-leaved / coniferous ]), an analysis of different grid densities and the analysis of the option to use differing grid 
densities in different strata and a comparison with other stratification techniques for forest inventories for forest 
enterprises as developed by Bóckman et al. (1998). 
4 USING THE KNN TECHNIQUE 
4.1 Requirements and data preparation 
A large number of sample plot data that are geo-referenced and satellite data also geo-referenced and topographically 
normalised are a prerequisite for applying the k-nearest-neighbour method (Tompoo & Pekkarinen, 1997). The first 
analysis was made on base of the Landsat TM data; the analysis based on IRS 1C LISS data is under preparation. The 
Landsat TM 5 data have been supplied in an already geo-referenced and topographically normalised form by the forest 
research institution of the state Nordrhein-Westfalen LOBF, Miinster. Details on the data processing are given in 
Diemer & Lucaschewski (1999). The TM-channels 1 to 5 and 7 have been used. All of the 817 sample plots have been 
used, both the 735 that belonged to the state forest and 82 plots that have been assessed in the forest of a public 
foundation. 
4.2 Methods 
The k-nearest-neighbour method for quantitative attributes postulates that there is a context between a measurable 
physical attribute, such as the timber volume and the spectral signature of multi-spectral remote sensing data. If for a 
large number of sample points ("reference points") the spectral values of corresponding pixels and terrestrial 
measurements of attributes are available, an estimate can be determined for all pixels for which no information from 
sample data is available using a simple method that does without model assumptions: for every pixel s, the Euclidean 
distance to all reference points v= I..n (i.e., to their corresponding pixels) is determined as a measurement for the 
similarity of the signature (Tompoo & Pekkarinen, 1997): 
| I 
= S. -B,y (9) 
i=1 
E. Euclidean distance of pixel s to reference point v 
I: number of channels 
D value of the reference point v in channel i 
Dis value of the pixel s in channel i 
The reference points k (j=1..k) with the closest distance are then selected from all n reference points. A weight W, is 
assigned to each of the selected k reference points so that the sum of all k weights is 1 and the weight is reversely 
proportional to the square of the Euclidean distance is (Tompoo & Pekkarinen, 1997): 
UE? 
W eU ET (10) 
SUME 
j= 
Let the measurement of the attribute on the reference point j be Mj. The estimate value M , for pixel s is then calculated 
as the balanced mean of k values (Tompoo & Pekkarinen, 1997): 
M, -Xv W,*M, (11) 
  
360 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000.