Colorado/Tom and Miller 
generate the data set. The boundaries of the sub-image were adjusted to the USGS quadrangle 
map with the use of an MSS band 7 graymap. This geometric rectification allowed the spatial 
registration of Landsat data with ancillary elevation, slope, aspect, and airphoto-derived 
vegetation data (Fig. 6.2). A fifth ancillary variable, Landsat image insolation, was calculated 
from the slope and aspect variables, and other information, to yield the incident solar radiation 
on the terrain at the time of the Landsat overpass. Thus, the original four MSS bands were 
registered with five ancillary map variables and overlayed as 1 ha. cells. 
Stepwise linear discriminant analysis was applied to these nine variables using the site indeces 
previously measured at the 37 training sets as ‘ground truth". The classification of these in- 
dividual test plots using the four Landsat MSS bands and the five ancillary variables yielded 
34 correctly classified plots out of 37, for an average training set accuracy of 92% (Fig. 6.2). 
A reclassification was performed to determine the training set accuracy which could be 
achieved using only the five map-derived variables (elevation, slope, aspect, vegetation and 
insolation). The resultant accuracy was 68% (Fig. 6.3). Using a similar approach, the opti- 
mum combination of five variables was selected from the total of nine available. The five 
optimum variables in order of decreasing importance were elevation, MSS band 6, slope, MSS 
band 4, and vegetative cover, and they yielded a training set accuracy of 81% (Fig. 6.3). Thus, 
the substitution of two easily obtained Landsat variables (MSS bands 4 and 6), for two tedi- 
ously obtained map-derived parameters (aspect and insolation), increased the accuracy by 
an absolute 13%, which represents 41% of the remaining improvement possible in the range 
of 68% to 100%. 
These results indicate that easily obtainable Landsat variables can be used to significantly im- 
prove the accuracy of predicting forest site index and potential productivity. 
Further Information: 
References 
Getter, J. R. and C. Tom. 1977. Forest site index mapping and yield model inputs 
to determine potential site productivity. Colorado State University, Colorado 
State Forest Service, Preliminary Report, Ft. Collins, Colorado. 20 p. 
Experimenters 
Craig Tom, HRB-Singer, Inc., Environmental Analysis Group, Science Park, State 
College, Pennsylvania 16801 U.S.A. 
Lee D. Miller, Texas A&M University, Remote Sensing Center, College Station, Texas 
77843 U.S.A.