IEEE Geoscience and Remote Sensing Magazine - September 2021 - 60

RGB values of the Potsdam data set are already provided as
8-b values, the quantization parameter is not assessed on
Gray-Potsdam.
In Table 12, we present the classification results of the
Pavia data set with 8- (default), 16-, and 64-b quantization.
Regarding the classification based solely on the four PCs,
approximating the 64-b float values to 16-b values had
little influence on the results. However, approximating the
64-/16-b to 8-b values had a larger impact on the classification
results, with a decrease of more than 3% in terms of
accuracy and l scores.
A more significant loss is observed in the classification
results with AP-minT and AP, with a decrease of more than
5% in terms of accuracy and l scores when the 64-b float
values are approximated to 8-b values. On the other hand,
we have opposite results for SDAP,
a -AP, and ~ -AP: using
8-b for quantization provides the best scores. Consequently,
we could draw different conclusions regarding the
performance of AP compared to SDAP,
a -AP, and ~ -AP
depending on the quantization settings: if 8 b, then SDAP,
a -AP, and
forms those three methods.
Regarding FPs, the best FPn
and FPa
and FP an+
achieved with 64-b quantization, whereas the best FPa
results are
is
achieved with 8-b quantization. It is noteworthy that, as
discussed previously, FPn
also presented opposite
results with respect to connectivity parameters. This may
suggest that having finer (respectively, coarser) components
trees obtained with four-connectivity and 64-b quantization
(respectively, eight-connectivity and 8-b quantization),
for example, leads to better FPn
(respectively, FP )a
scores.
Our experiments show that, in addition to the choice of
attributes and threshold values, the connectivity and quantization
parameters have a great impact as well on the performance
of APs with different tree representations. More
importantly, this effect varies greatly depending on the underlying
tree type. In summary, these two parameters, which
are almost always silently set to default values (unmentioned
in published studies), appear to merit the same level of attention
and care that threshold/attribute selection enjoys.
DISCUSSION OF THE GENERALIZATION OF APs
The experiments described in the previous section follow
the same approach of other experiments in the literature:
the training and testing features are obtained from the
same tree computed on the whole data (or on the PCs of
the data). This approach is reasonable when the aim is to
completely classify the pixels of an image whose annotated
pixels are evenly spread across this image [12]. That
was the case of the training sets considered previously,
which allowed us to obtain an improvement of more than
30% in terms of classification accuracy using APs with
respect to spectral pixel values. However, this technique
raises doubts in situations where one encounters distinct
images for training and testing or, alternatively, when the
training pixels are not evenly spread across the data. This
60
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE SEPTEMBER 2021
~ -AP outperform AP; otherwise, AP outperTABLE
12. A COMPARISON OF THE CLASSIFICATION RESULTS OF PAVIA FOR DIFFERENT QUANTIZATION PARAMETERS.
METHOD
Four PCs
AP-maxT
AP-minT
AP
SDAP
a-AP
~-AP
FPn
FPa
FPn + a
DIMENSION
4
80
80
152
80
80
80
152
152
304
OA (%)
65 ..
27 025!
89.21 ! 0.63
87 .11 2!
90 ..
57 26!
94.1 ! 0.21
95.46 ! 0.59
96.33 ! 0.4
96 ..
76 015!
85.48 ! 0.98
96 ..
15 014!
CLASSIFICATION RESULT (8 B)
AA (%)
74 ..
87..
92 ..
93 ..
88 02!
37 019!
62 061!
13 067!
93.85 ! 0.37
95.25 ! 1.23
97.17 ! 0.89
97 ..
05 019!
93.94 ! 0.55
96 ..
74 032!
l # 100
56 ..
93 027!
85.42 ! 0.8
83 ..
87 ..
23 241!
53 332!
92.16 ! 0.27
93.91 ! 0.79
95.08 ! 0.54
95 ..
65 021!
81.36 ! 1.16
94 ..
82 019!
OA (%)
68.6 ! 0.1
88..
91 ..
95 ..
92..
93..
93..
97 ..
82..
97 ..
The values in bold represent the highest OA among the scores achieved using the three different quantization settings.
CLASSIFICATION RESULT (16 B)
AA (%)
97 02!
58 164!
43 021!
29 032!
87 072!
75 054!
09 07!
72 074!
48 031!
78.08 ! 0.09
87 ..
94 ..
94 ..
91 ..
93 ..
92 ..
97 ..
92 ..
39 015!
55 065!
01 047!
36 042!
85 022!
97 076!
43 047!
64 063!
98.3 ! 0.19
l # 100
60.93 ! 0.11
84 ..
88 ..
93 ..
89 ..
91 ..
91 ..
96 ..
99 027!
89 206!
83 028!
66 042!
82 093!
64 071!
1093!
77 ..
96..
99 095!
62 041!
OA (%)
CLASSIFICATION RESULT (64 B)
AA (%)
68..
89 ..
51 024!
07 037!
92.42 ! 1.07
95.5 ! 0.19
92 ..
94 ..
93 ..
97.10 ! 1
83 ..
49 047!
08 017!
75 044!
09 042!
97.55 ! 0.29
78.08 ! 0.22
87.42 ! 0.17
94.77 ! 0.83
94.25 ! 0.52
91 ..
93 ..
92 ..
34 061!
96 009!
78 098!
97.64 ! 0.33
93 ..
98..
05 029!
23 026!
l # 100
60..
85..
81 029!
13 049!
89.97 ! 1.35
93.93 ! 0.26
89 ..
92 ..
92 061!
1022!
91 ..
64 059!
96.11 ! 1.31
78 ..
5055!
96.71 ! 0.39
IEEE Geoscience and Remote Sensing Magazine - September 2021

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