Geoscience Reference
In-Depth Information
data in rows and classified map in columns. For instance, 28 reference units labeled
as forest have been classified as forest, whereas 26 have been mis-classified as non-
forest. This means only half of the reference pixels have been classified correctly
(producer accuracy is 52%). The producer accuracy reflects the omission errors of
forest (reference units not found). The user accuracy reflects the commission errors
(e.g. pixels mapped as forest that are not labeled as such in the reference data). The
accuracy is consistently higher for non-forest than for forest. The numbers reported
in brackets represent a normal approximation of the 95% binomial proportion con-
(Congalton 1991) is rather low (0
.
56), which shows the need for further improvement
with, e.g. a supervised classifier.
From the assessment of the supervised map, it is shown that the accuracy has
indeed improved. The accuracy is consistently higher than for the unsupervised map
for both the forest and non-forest classes.
pkdiff -ln valid -ref reference.sqlite -i LC82070232013_fmap_supervised_
ₒ
masked.tif -cm -nodata 0
-lr label -lc class -o reference_fmap_
ₒ
supervised.sqlite
processing layer valid
0...10...20...30...40...50...60...70...80...90...100 - done.
12
1 4 0
2
10
846
class #samples userAcc
prodAcc
1
54
81 (12.4) 81 (5.3)
2
856
99 (3.4)
99 (0.4)
Kappa: 0.80
Overall Accuracy: 98.0 (0.47)
To analyze the effect of the border pixels on the accuracy assessment, we just
have to access the layer named as “border”. To include all pixels (valid and border),
we calculate the confusion matrix and accuracy measures on both layers. The overall
accuracy for the forest map based on the supervised classifier has dropped from 98
to 94% and the Kappa value from 0
65. For the unsupervised classifier, the
overall accuracy dropped from 94 to 91% and the Kappa value from 0
.
80 to 0
.
.
.
42.
The output vector dataset from
pkdiff
10
contains an attribute for the reference
label (named
label
) and for the classified map (named
class
). By comparing them, we
can identify on amapwhich pixels have been correctly classified andwhich have been
misclassified. Pixels labeled as forest, but classified as non-forest are referred to as
omission errors
(of forest). Pixels incorrectly classified as non-forest are referred
to as
commission errors
(of forest). We put this information in a new attribute
assessment
of the vector
reference_fmap_supervised.sqlite
and
visualize the spatial distribution of the classification errors inQGIS. This is illustrated
in Fig.
17.7
.
48 to 0
10
reference_fmap_supervised.sqlite
.
Search WWH ::
Custom Search