VII. ACCURACY ASSESSMENT
To perform the accuracy assessment, the error matrix is
utilized. ˆK values (Kappa coefficient) > 0.80 represents strong
agreement or accuracy between the classification map and the
ground reference information. ˆK values between 0.40 and 0.80
represents for moderate agreement. ˆK values < 0.40 stands for
poor agreement [20] .There are two categories of ground
reference, one is the various categories with their spatial
distribution maps being determined before classification, and
the other is the existing local land-cover map of the area [22].
Cross validation was used to do the accuracy assessment.
ENVI plug-in Toolbox ---ROI separator is utilized, with
randomly selected 67% ROI as the training sample, and the rest
(33%, see the Ground Truth in table III.) to validate. The rest
(33% ROI) would not act as the training samples of
classification [23].
Shadow and cloud in initial classification did not appear in
ultimate classification results. Cropland and urban &rural are
divided into several land-cover types in the final classification.
Therefore, only the remaining unaltered four classes appear in
the confusion matrix (table III).
The overall classification accuracy is 91.50%. The producer
accuracy of bareland is 100% and the user accuracy of bareland
is 51.81%, indicating that there exists no omission but
commission error in bareland classification. The producer
accuracy of grass is 64.29%. 21.43% of grass was misclassified
into woodland, and 14.29% of grass was misclassified into
bareland. The producer accuracy of woodland is 86.48% and
the user accuracy of woodland is 98.21%. 1.57% of woodland
was misclassified into grass, and 11.95% of woodland was
misclassified into bareland. The producer accuracy of water is