Geoscience Reference
In-Depth Information
f
f
f
f
kk
kk
kk
PA
==
=
+
(15.9)
k
n
n
ÂÂ
1
+
k
f
f
f
ik
kk
ik
i
=
i
ik
=
π
1
By substituting Equation 15.8 and Equation 15.9 into Equation 15.7, it is easily derived that:
Ê
Á
ˆ
˜
¥
1
1
REA
=
-
100
(15.10)
k
UA
PA
k
k
Thus, REA can be obtained using information on the error matrix or the user's and producer's
accuracy.
Under the assumption that the distribution of errors in the error matrix is representative of the
types of misclassification made in the entire area classified, it is easy to calibrate area estimates
with REA or UA and PA as follows:
Ê
Á
ˆ
˜
¥
f
N
f
N
1
1
kk
kk
AA
=
-
¥
REA
=
A
-
¥
-
100
(15.11)
c k
,
pc k
,
k
pc k
,
A
A
k
k
where
A
= calibrated area in percentage for a given land cover type
k
and
A
= precalibrated
c,k
pc,k
area in percentage for a given land cover type
k
.
15.3 RESULTS
Figure 15.1 shows the means and standard deviations of nine landscape indices for three
accuracy groups. Except for PD and MPS, landscape indices had < 10% differences in their means
among three accuracy groups. The standard deviations of the landscape indices in the lowest
accuracy group are much higher than those in the higher accuracy groups. The differences in
standard deviations between the lowest accuracy group and other two accuracy groups exceeded
100%, indicating that the uncertainties were higher when classification accuracy was lower.
The statistics of classification accuracy, including the overall accuracy, producer's accuracy,
and user's accuracy, all have differences of < 20% among the three accuracy groups (Figure 15.2a).
The standard deviation values for overall accuracy are also about the same among the three accuracy
groups but are clearly different for producer's accuracy and user's accuracy (Figure 15.2b). Maps
in the lowest accuracy group have much higher variations in producer's accuracy and user's accuracy
than those in the other two accuracy groups.
For a few indices, such as MPDF, AWMPFD, and SDI at the landscape level, no matter what
the classification accuracy was, the errors of landscape indices were within a range of 10% (Figure
15.3). If classification accuracy was poor, the errors of some other landscape indices exceeded
100%. They include PD, PSCOV, ED, AWMSI, and MPI for entire landscapes or forest patches
(Figure 15.3 and Figure 15.4). Although no constant relationships were found between the overall
accuracy and landscape indices, maps with higher classification accuracy resulted in lower errors
for most landscape indices (Figure 15.3 and Figure 15.4). However, overall accuracy did not have
good control over the variations of landscape index errors and therefore was not a reliable predictor
for the errors of landscape indices. This was particularly true when the overall accuracy was
relatively low.
