to the cloud and water image. These classes, being of no interestto forest clearing and regrowth analysis, were masked for all
dates of imagery to avoid confusion in the change-detectionclassification.
is the mean
for the bright target of the referenceimage, Bs is the mean
for the bright target of the subjectimage, Dr is the mean
for the dark target of the referenceimage, and
s the mean
for the dark target of the subjectimage. Using linear regression, the corrected pixel values forthe subject image
were calculated from the original
for each band
by the following equation:
Three change-detection methods (NDVIifferencing, PCA, and
classification) were independently applied to thecloudlwater-masked and radiometrically normalized time-series
data set. A three-date forest change-detection classi-fication of the selected study area was generated from eachmethod. Each method was evaluated and compared with theother methods on its ability to classify temporal states in forestcover(i.e., cleared, regrown, no change) over the three timeperiods. The methods were evaluated and contrasted on thebasis of classification accuracy (Congalton,1991),efficiency ncomputation and processing, and ease in interpretation.NDVI Image DifferencingDifference images were created by first calculating NDvI valuesfor each date (j)of imagery by the following equatioa:Principal ComponentAnalysisThe principal component transformation was performed sepa-rately on two data sets (1993 and 1995,1995 and 1997)usingthree
bands (3,4, and 5) for each date. Each two-date data setcontained six bands. The transformation used the "prince"routine (ERDAS 1997), modified to calculate the transformfrom a correlation matrix of the data set. Several authors havecompared this "standardized" approach to
against transfor-mations based on the covariance matrix (Conese et al., 1988;Eastman and
1993; Rencher 1995). Reported advantagesof the standardized approach include improved interpretabil-ity, the isolation of seasonal effects and variability due to noise,better statistical control, and more precise classification. Foreach data set, the "standardized"
routine output includedsix component images, a table of eigenvalues quantifying theproportion of variance explained by each component, and amatrix of eigenvectors (weights or factor loadings) depictingbetween-date correlation for each band with each component.Components that represent change typically show an ibsenceof correlation amone bands between dates IBvrne et al.. 19801.The component that'best highlights the chGge of interest is
chosen for thresholding, using visual interpretation of compo-nent images and analysis of the eigenvector matrix.Image interpretation was based on the assessment of spa-tial continuity, by seeking out the components that express thedifferences in the changes of interest as spatially discontinuousareas within the image. The eigenvector analysis examined thealgebraic signs on the weights. Differences between dates areexpressed by the weight of one band at one date having anopposite sign to that of the same band of the other date. Basedon these criteria, two of the six components (components3 and4 for each two-date data set) were selected from the
forthresholding of no-change and change areas.
these two com-ponents, the one that showed the highest ability to thresholdforest
i.e., the highest estimatedKappa according to the reference sample points) was chosen forfinal classification.