PE&RS July 2016 Public - page 530

Change Detection Results
The multi-spectral mean values of the corresponding patches
in bi-temporal images are used to calculate two change crite-
ria: Image Differencing and
MAD
Transform (Nielsen, 2011).
The final product of each criterion for change detection is a
sort of binary classification: changed or unchanged. However,
since we want to prove if the comparison of the spectral prop-
erties of the corresponding patches are suitable for change
detection, we generated the
ROC
(Receiver Operating Charac-
teristic) curves of the two different change criteria for the four
study datasets.
To evaluate the effectiveness of each criterion in change
detection, we compare the change results to a reference data
(Congalton and Green, 2009), which are generated through
manual inspection. In each dataset, a group of around 100 un-
changed and changed reference patches are selected and their
change status are checked against the results generated using
each of the two change criteria shows samples of reference
patches (Figure 11).
The mentioned process of checking the labels of the refer-
ence patches is repeated for a fine range of thresholds as-
sociated with each change criterion. For every threshold the
values: true positive (tp), true negative (tn), false positive (fp),
and false negative (fn) are counted. True positive and true
negative are related to the patches which are truly identified
as changed and unchanged, respectively; false positive and
false negative are related to the patches which are falsely
identified as changed and unchanged, respectively.
Using the mentioned values, a
ROC
curve, which is a plot of
True Positive Rate (
TPR
)
versus
False Positive Rate (
FPR
)
(Equa-
tions 9 and 10), for each criterion across different thresholds
is generated (Fawcett, 2006):
Figure 9. Comparison between the coregistration results of (a) the conventional method and (b) the PWCR method. (An enlarged view).
Figure 10. Comparison of the Area Ratios generated using the PWCR method and the conventional method for datasets DT1 to DT4.
530
July 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
447...,520,521,522,523,524,525,526,527,528,529 531,532,533,534,535,536,537,538,539,540,...582
Powered by FlippingBook