selection of either the training or test fields. Two common measures 
of the accuracy of classification maps are the percent correct clas- 
sification within test fields and overall correct classification.  Un- 
less sound statistical sampling procedures are used to select test 
fields, such numbers become quite suspect and may or may not represent 
the actual accuracy of the map. Such accuracy measures cannot be used 
to establish sampling errors or confidence statements relative to a 
specific resource estimate. In some cases, these results will be use- 
ful to scientists in the remote sensing community for evaluating the 
performance of classifiers, analysis procedures, etc. However, such 
statistics have little practical value to a resource manager requiring 
an estimate of the number of hectares in a resource class with a con- 
fidence interval of plus or minus 5% at the .95 probability level. 
A more thorough analysis of the accuracy of classification maps is 
needed as are sound statistical sampling procedures to achieve estimates 
around which a confidence statement can be placed. It is interesting 
to note that there exists a heavy reliance on statistical decision theory 
in classifying multisp jectrat data but an apparent reluctance to use 
statistical sampling theory to evaluate the results and provide estimates 
in a format required by the user community.  Hord and Brooner (1976) 
present a methodology for estimating the classification accuracy, place- 
ment of type lines, and geometric accuracies. Nichols et al (1974) and 
Langley (1975) have described applications of LANDSAT multispectral data 
and multistage sampling theory to derive statistical estimates of re- 
Source parameters with estimates of sampling error and confidence inter- 
vals. It is incumbent on the remote sensing community to make an inten- 
sive effort to develop sound sampling procedures to evaluate the accuracy 
of map products and generate statistical estimates of resource parameters 
with specific confidence statements based on user objectives. Map pro- 
ducts and tabular data presented in such a manner will gain wider user 
acce aptance and be in a format which allows a resource manager to make a 
de ion on the value of the data and incorporate such data into his 
or 
u 
management decis sions. 
5 
Geometric correction, image registration 
Although considerable geometric distortions are currently present 
in LANDSAT film and CCT products, algorithms have been developed which 
will produce ected images with a nominal RMS error of 
y 
e 
  
approximately 60 meters. A major task in performing geometric correction 
of LANDSAT data is the accurate location of ground control points. These 
control points must be digitized with respect to their geographic lo- 
cation and an accurate determination of their line and column coordinates 
c 
in the LANDSAT data must be made. Presently, a large number of invest- 
igators have independently developed geometric correction algorithms and 
have established sets of ground control points for specific areas of in- 
r 
terest. The establishment of a data bank of ground control points, avail- 
able to all users, would reduce the effort required to digitize the geo- 
graphic coordinates of such points. Further, as new control points are 
established, the data bank can be updated and the availability of more 
control points will theoretically lead to more consistent geometric cor- 
rection results. 
It is intended that future satellite data, including both film and