411 
After the process of reformatting and geometric correction was completed, 
three techniques for extracting temporal information were applied to an 
initial test site and compared to the FA technique. The best technique 
was determined visually, based on selected ground-truth features. The 
FA technique was also applied to the four additional test sites and 
evaluated. These four test sites were analyzed visually for correct 
classification of temporal information, based on interpretation of the 
original aerial photos. The three techniques to be evaluated were: 
1. A classification map of the early, one-channel scene (using den 
sity slicing) and a classification of the late scene (using standard 
supervised and unsupervised classification techniques) were developed 
separately. The two classified scenes were then compared digitally 
after classification. The temporal information should appear as differ 
ences in cover types between the two scenes, which should be evident on 
the comparison map. 
2. The data from the early scene and the late scene were combined 
into a four-channel data set. A classification map was produced using 
normal supervised and unsupervised classification techniques. Those 
areas of high temporal information content should have unique spectral 
information and should map out as unique features or cover types. 
3. A principal component analysis was performed on the covariance 
matrix of the late scene. Each principal component was scaled and 
translated over a range from 0 to 255. Similarly, the black-and-white 
scene was scaled and translated from 0 to 255. Differences and ratios 
between the early scene and each principal component were calculated, 
density sliced, and displayed. The areas of high-temporal information 
content should appear as extreme values in the ratio and difference data. 
These three techniques were then compared to an FA technique applied to 
the same test site. The basic idea in this technique was to find a 
unique set of underlying factors in the combined data set and rotate 
these factors to find an axis, or transformation, that highlights the 
contrasts between the early and late scenes. 
The FA technique starts with a correlation matrix derived from the four- 
channel data set of the test site. The correlation matrix was input 
to the SPSS factor analysis package (Nie et al., 1975) for factor rota 
tion. Five rotational methods were considered: principal components, 
equimax, varimax, quartimax, and oblimin. Although the principal com 
ponents are the basic, unique set of initial factors for the other 
methods, they were examined because they may also contain a contrast of 
interest. The remaining rotations started with the principal components 
but then rotated the principal axes to highlight other aspects of the 
data set. The resulting transformation matrix and correlation matrix 
(between the rotated factors and the original channels) were examined 
for factors that contrast the early and late scenes and that were 
relatively uncorrelated to the original channels. 
The most promising transformation was used to transform the original 
data. Initially, the selected transformation matrix was adjusted to 
account for the fact that the original data were not standardized. Each 
factor in the matrix was multiplied by a diagonal matrix of inverse 
standard deviations to account for data standardization. The data were 
transformed using this adjusted transformation matrix, scaled and trans 
lated, and density sliced and displayed. The extremes in the trans 
formed data contain the most temporal information.