ratio
among the three methods, which caused its curves to
be positioned in the upper left, as shown in Figure 5a, 5c,
5e, and 5g. The left-shift of the
LBLW
curves was smaller than
that of the
LBLW
&DS curves, but exhibited an obvious up-shift
compared with the LW curves. This finding denotes that the
LBLW
method exhibited a slight decrease in
precision
and an
obvious increase in
recall ratio
compared with the LW meth-
od. Furthermore, if the
precision
measure was fixed, then the
recall ratio
s of the three methods were ranked as
LBLW
&DS
>
LBLW
> LW. Meanwhile, if the
recall ratio
was fixed, then
the
precision
measure was ranked in the same order. Analy-
ses of the
recall-precision
curves also showed that both the
LBLW
&DS and
LBLW
methods were significantly superior to the
LW method. The conclusions coincide with the results of the
analyses of the
F-measure
curves.
In Table 2,
F-measure
was the highest when the
LBLW
threshold was 3.0. The corresponding classification results were
selected for further visual interpretation and analysis. In the
two experimental areas, impervious surfaces were extracted
totally although several misclassifications with other classes
were included. For example, the sands (bare lands) in the up-
per left corner of Plate 2a were misclassified as an impervious
surface because they produced similar spectra as an impervious
surface. First-level classification errors were also transmitted to
second-level road extraction. Several false impervious surfaces
in this corner were misclassified as roads by the LW and
LBLW
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
Figure 5. Accuracy analyses of the two experimental areas. (a) and (e)
Precision–recall
curves of areas 1 and 2, respectively, obtained by
counting region number; (b) and (f)
F-measure
curves of areas 1 and 2, respectively, obtained by counting region number; (c) and (g)
Pre-
cision–recall
curves of areas 1 and 2, respectively, obtained by counting region area; and (d) and (h)
F-measure
curves of areas 1 and 2,
respectively, obtained by counting region area.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
February 2016
155
