(scale 20). Analyses on measure
m
2
by same-scale comparisons
also indicate an improvement in method accuracy.
Plates 1 to 3 show several segmentation results of the refined
and original methods for the experimental areas. The results were
obtained with the best inputs presented in Table 2. The perfor-
mance of the methods were then further evaluated and compared.
In
HBC-SEG
, most of the segments grew to quasi-object scales
and maintained low under-segmentation errors and high
boundary precision under small-scale parameters (10 to 20).
However, parts of the objects (e.g., buildings) still exhibited
possibly unfavorable over-segmentation errors probably
because of the weak edges extracted and trivial inner details.
Such deficiency would influence the succeeding
OBIA
steps if
the goal is building extraction. The best scheme is to solve this
under-segmentation error within the segmentation framework.
From scales 10 to 20, the OPs in the original and refined
methods were merged but at different degrees (Plates 1, 2, and
3). The refined method performed more merging instances
without inducing under-segmentation errors. After refine-
ment, the instance of re-merging of segments was fewer in
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
(i)
Plate 1. Segmentations of HBC-SEG and the refined method on Area 1: (a) the original data, (b) the reference map, and (c) straight line extracted
by Burn’s method. Plates 1d and 1e are the segmentation results obtained with HBC-SEG at scales 10 and 20, respectively. The red boxes in
Plate 1e denote several (not all) distinguishable over-segmentation errors for HBC-SEG at scale 20, which were resolved by the refined method
at scale 10. Plates 1f and 1g are the segmentation results obtained with HBC-SEG at scales 40 and 50; the red boxes denote several (not all)
under-segmentation errors. Plates 1h and 1i are the segmentation results obtained with the refined method at scales 10 and 20, respectively.
T
able
2. M
ethod
A
ccuracy
A
nalyses
. I
n
the
T
wo
A
reas
, R
ecall
R
atio
r
E
xhibits
an
A
verage
I
ncrease
of
0.048 (S
cale
10, A
rea
1), 0.042 (S
cale
20, A
rea
1),
0.041 (S
cale
10, A
rea
2),
and
0.036 (S
cale
20, A
rea
2); A
long with
V
ery
S
mall
R
eductions
in
P
recision
p
that
are
L
ess
than
0.006
Input
HBC-SEG
Refined method
Area 1
Area 2
Area 1
Area 2
Scale
w p
r
m
2
p
r
m
2
p
r
m
2
p
r
m
2
10
1.0 0.857 0.423 0.373 0.852 0.446 0.388 0.847 0.470 0.408 0.852 0.479 0.414
0.9 0.854 0.424 0.371 0.837 0.441 0.374 0.852 0.467 0.408 0.834 0.480 0.404
0.7 0.856 0.416 0.367 0.842 0.414 0.359 0.851 0.471 0.409 0.835 0.467 0.398
Average 0.856 0.421 0.370 0.844 0.434 0.374 0.850 0.469 0.408 0.840 0.475 0.405
20
1.0 0.845 0.464 0.402 0.830 0.552 0.461 0.836 0.500 0.427 0.843 0.560 0.480
0.9 0.842 0.471 0.410 0.825 0.538 0.431 0.839 0.511 0.442 0.815 0.558 0.450
0.7 0.844 0.472 0.410 0.821 0.492 0.404 0.839 0.512 0.439 0.819 0.571 0.478
Average 0.844 0.469 0.407 0.825 0.527 0.432 0.838 0.511 0.436 0.825 0.563 0.469
402
May 2015
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
