IEEE Geoscience and Remote Sensing Magazine - September 2021 - 62

GENERALIZATION OF APs TO OTHER
PARTIALLY ANNOTATED IMAGES
The " standard " training set of the Pavia University data set,
i.e., the training set of
Pavi ,a1 is composed of pixels belonging
to most of the ground-truth (and testing) CCs. Hence,
the training set of each semantic class accounts for most of
the variability in terms of spectral signatures and geometric
properties of the CCs of said class. However, the availability
of training samples from every CC of each semantic class
in a real-world scenario, dealing commonly with remote
sensing data sets representing geographically large areas, is
evidently unrealistic. The data-splitting procedure used in
this section aims to simulate, to a limited degree, the aforementioned
real-world conditions.
We compare the classification results presented in the
previous section with those based on a new partition of
the Pavia and Gray-Potsdam data sets. (The proposed split
of the Pavia and Gray-Potsdam data sets is available at
https://gitlab.inria.fr/dsantana/attributes-profiles-survey
-source-codes.) In the new partition of Pavia [see Figure 8(b)
and (c)], 10 sets of training samples were randomly extracted
from a restricted region (composed of 85 CCs) of the ground
truth, which is composed of 229 CCs in total. To provide
comparable results with the standard partition of this data
set, we selected the same number of training samples per
class, resulting in 3,921 training samples (see Table 1) and
40,002 testing samples for each of the 10 random splits.
Similarly, 10 sets of training samples of Gray-Potsdam
were extracted from nearly half of the ground-truth CCs. In
total, 726 ground-truth CCs (see Figure 9) contributed to
the new training sets of Gray-Potsdam. The same numbers
of training and test samples given in Table 2 were obtained.
These new partitions of the Pavia and Gray-Potsdam data
sets are denoted as Pavia2
and GrayPotsda ,m2
respectively.
Both the training and testing sets include pixels belonging
to all classes of those data sets. The experiments were
performed using the default connectivity and quantization
settings given in the " Experimental Study " section which
are four-connectivity and 8-b quantization. We applied the
same set of thresholds for the area and moment of inertia
attributes used in the previous experiments on Gray-
Potsdam and Pavia.
The main challenge of using those new data set partitions
is that a sample of a given semantic class is not always
representative of other samples of the same class in different
regions of the image. This is particularly true for the bare soil
and meadows classes of Pavia. In the classification results
based only on pixel intensities, several meadow samples are
classified as bare soil and vice versa, as shown in Figure 5(d).
A similar observation can be drawn for the building and impervious
surface classes of Gray-Potsdam [see Figure 6(c)].
Tables 13 and 14 present the classification results computed
on the partitions Pavia1
(Figure 3) and Pavia2
(a)
(b)
FIGURE 9. (a) The original Gray-Potsdam ground truth. (b) The CCs
that contributed to the training set of the partition GrayPotsdam2.
(Figure 8).
First, in Table 13, we see a degradation in the baseline classification
results (four PCs) using the new partition, which
is due to the training set not including diversified samples
of every class. Then, we can observe an even larger degradation
in terms of AP results. With Pavia1, all APs improve the
TABLE 13. THE CLASSIFICATION RESULTS OF THE PAVIA DATA SET PERFORMED ON THE TRAINING/TESTING SETS
FROM STANDARD PARTITION PAVIA1 (FIGURE 3) AND NEW PARTITION PAVIA2 (FIGURE 8).
METHOD
Four PCs
AP-maxT
AP-minT
AP
SDAP
a-AP
~-AP
FPn
FPa
FPn + a
TF-AP
DIMENSIONS
4
80
80
152
80
80
80
152
152
304
72
OA (%)
CLASSIFICATION RESULT (PAVIA1)
AA (%)
65..
89 ..
87 .11 2!
90 ..
94..
27 025!
21 063!
57 26!
1021!
95 ..
96 ..
46 059!
33 04!
96.76 ! 0.15
85 ..
96 ..
93 ..
48 098!
15 014!
98 053!
The values in bold represent the highest score in each column.
62
74 ..
87 ..
92 ..
93 ..
93 ..
95 ..
88 02!
37 019!
62 061!
13 067!
85 037!
25 123!
97.17 ! 0.89
97 ..
93 ..
96 ..
96 ..
05 019!
94 055!
74 032!
26 029!
l # 100
56 ..
85 ..
83 ..
87 ..
92 ..
93 ..
95..
93 027!
42 08!
23 241!
53 332!
16 027!
91 079!
08 054!
95.65 ! 0.21
81 ..
94 ..
92 ..
36 116!
82 019!
04 068!
OA (%)
59 ..
61 ..
68 ..
64 ..
76 ..
64 ..
63 ..
66 ..
CLASSIFICATION RESULT (PAVIA2)
AA (%)
58 074!
23 068!
72 272!
49 135!
63 37!
72 089!
62 15!
75.57 ! 1.43
4117!
72 ..
69 ..
06 291!
25 56!
73 ..
82 ..
79 ..
84 .074!
82 ..
76 ..
74 ..
44 045!
27 036!
89 106!
05 108!
02 119!
59 078!
85.36 ! 2.14
81 ..
82 ..
84 ..
48 039!
83 083!
46 18!
l # 100
50 ..
54 ..
62 ..
57 ..
57 ..
55 ..
69..
59 ..
65 ..
63 ..
76 076!
42 069!
03 288!
9138!
70.85 ! 4.09
32 111!
96 155!
61 163!
8117!
81 315!
22 594!
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE SEPTEMBER 2021
https://gitlab.inria.fr/dsantana/attributes-profiles-survey-source-codes https://gitlab.inria.fr/dsantana/attributes-profiles-survey-source-codes
IEEE Geoscience and Remote Sensing Magazine - September 2021

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