IEEE Geoscience and Remote Sensing Magazine - September 2021 - 64

region are fairly well classified when using the partition
GrayPotsda ,m1
as shown in Figure 6.
Overall, using the new partitions Pavia2
and GrayPotsda ,m2
we observed a very significant drop of performance across all
approaches. Moreover, the drop in performance occurred at
different degrees for each thematic class of the data sets. This
may be linked to the fact that some classes are composed of
regions with similar geometric properties (e.g., the asphalt
class of Pavia and the car class of Gray-Potsdam), while this
is not true for the other classes (e.g., the meadow class of Pavia
and the low vegetation class of Gray-Potsdam). However, a
deeper investigation is necessary to confirm this assumption.
GENERALIZATION OF APs TO
MULTIPLE-IMAGE DATA SETS
Though the partition Pavia2
pixels of Pavia2
may still share features obtained from nodes
at higher levels of the tree. The same holds for the partition
GrayPotsda .m2
In this section, we go one step further to completely
separate the computation of training and test features.
As mentioned previously, the Potsdam data set is composed
of several patches covering a large urban scene. The
original International Society for Photogrammetry and Remote
Sensing (ISPRS) labeling contest consisted of providing
the classification labels for pixels on the testing patches
based on the information given by the training patches.
Hence, extracting training and testing features from the
same tree is possible only if all patches are connected in the
real scene. Otherwise, we would need to compute tree representations
separately on the training and testing patches.
To approach the generalization of APs to this kind of
test samples when compared to Pavi ,a1
better separates training and
the training and test
problem, we split the Gray-Potsdam image in such a way
that two independent component trees can be computed
TABLE 15. THE CLASSIFICATION RESULTS OBTAINED WITH THE PARTITIONS GRAYPOTSDAM1 AND GRAYPOTSDAM2
OF THE GRAY-POTSDAM DATA SET.
METHOD
Four PCs
AP-maxT
AP-minT
AP
SDAP
a-AP
~-AP
FPn
FPa
FPn + a
TF-AP
DIMENSIONS
1
20
20
38
20
20
20
38
38
76
18
OA (%)
CLASSIFICATION RESULT (GRAYPOTSDAM1)
AA (%)
l # 100
46..
63 ..
58 .023!
75..
76..
68 ..
68 ..
56 042!
65 014!
1005!
19 008!
32 007!
1006!
77..
77..
14 005!
19 005!
77.94 ! 0.04
72..
34 005!
34 ..
59 ..
56 ..
77 ..
78 ..
67 ..
67 ..
79 ..
79 ..
74 ..
3002!
68 002!
55 002!
56 002!
25 002!
94 004!
78 003!
41 002!
46 002!
80.17 ! 0.03
93 004!
The values in bold represent the highest accuracy scores in each column.
30 ..
52 ..
47 ..
67 ..
69 ..
59 ..
59 ..
70 ..
70 ..
84 032!
65 015!
13 021!
83 005!
2009!
47 008!
23 006!
42 006!
47 005!
71.43 ! 0.05
64 ..
33 006!
OA (%)
CLASSIFICATION RESULT (GRAYPOTSDAM2)
AA (%)
47 ..
52 ..
46 ..
57 ..
55 ..
76 029!
8019!
58 016!
11 019!
77 014!
58.31 ! 0.1
58 ..
57 ..
55 ..
56..
55 ..
17 013!
35 029!
53 029!
62 04!
52 009!
34 ..
45 ..
44 ..
55 ..
54 ..
53 ..
53 ..
54 ..
55 ..
54 ..
34 004!
39 009!
32 009!
54 012!
14 009!
23 021!
36 02!
56.16 ! 0.14
81 025!
68 019!
09 007!
l # 100
32 ..
39 ..
32 ..
44 ..
42 ..
15 021!
13 016!
97 014!
5023!
92 015!
46.38 ! 0.12
26 018!
87 034!
63 036!
9046!
75 009!
46 ..
44..
42 ..
43 ..
42 ..
TABLE 16. THE CLASSIFICATION RESULTS PER CLASS OF THE GRAY-POTSDAM DATA PERFORMED ON
THE PARTITION GRAYPOTSDAM2.
ACCURACY PER CLASS (%)
METHOD
Gray values
AP-maxT
AP-minT
AP
SDAP
a-AP
~-AP
FPn
FPa
FPn + a
TF-AP
DIMENSION
4
80
80
152
80
80
80
152
152
304
72
BACKGROUND
11..
33 ..
42 07!
17 058!
47.83 ! 1.01
98 032!
42 041!
68 142!
34 129!
48 021!
67 136!
53 062!
13 026!
45 ..
46..
35 ..
35 ..
46 ..
48 ..
47 ..
40 ..
TREES
15 ..
23 ..
64 146!
11 17!
28..
45 ..
41 ..
43 ..
43 ..
44 ..
46 ..
302!
43 ..
The values in bold represent the highest accuracy scores in each column.
64
56 068!
31 025!
47 028!
5029!
44 026!
47.49 ! 0.41
95 03!
52 036!
CARS
837026!
9042!
29 029!
78 025!
75 038!
38 035!
02 042!
..
47 ..
46 ..
77 ..
74 ..
63 ..
65 ..
79.3 ! 0.43
3038!
76 ..
78 ..
78 ..
34 044!
68 018!
BUILDINGS
48 ..
62 ..
61 ..
80 ..
77 ..
79 ..
78 ..
77 ..
79 ..
76 ..
LOW
VEGETATION
01 089!
83 06!
01 035!
02 034!
93 021!
05 017!
67 015!
80.08 ! 0.73
22 067!
18 099!
15 027!
73 ..
23 148!
73.86 ! 1.11
92 047!
7038!
66 035!
86 023!
82 043!
93 033!
63 03!
85 041!
54 028!
33 ..
37 ..
37 ..
37 ..
37 ..
36..
35 ..
35 ..
50 ..
IMPERVIOUS
SURFACES
49 ..
31..
48 ..
46 ..
46 ..
38 094!
505!
29 088!
44 025!
59 017!
59.91 ! 0.34
59 ..
46 ..
46 ..
46 ..
35 ..
88 036!
71 031!
08 033!
67 047!
73 021!
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE SEPTEMBER 2021
IEEE Geoscience and Remote Sensing Magazine - September 2021

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - September 2021
