Laser & Optoelectronics Progress, Volume. 61, Issue 4, 0437010(2024)
Multi-Scale Feature Extraction Method of Hyperspectral Image with Attention Mechanism
Figures & Tables(14)
Fig. 6. Structure of multiscale feature extraction method with attention mechanism
Fig. 7. Overall accuracy of proposed method by running 10 experiments independently with different hyperparameters. (a) Performance of SaAMN with different kernel sizes; (b) performance of SeAMN with different hidden sizes; (c) performance of SeAMN with different grouping strategies
Table 1. Categories and settings of Pavia University dataset
View tableTable 1. Categories and settings of Pavia University dataset
Class Class Name Training number Total number Total 900 42776 1 Asphalt 100 6631 2 Meadows 100 18649 3 Gravel 100 2099 4 Trees 100 3064 5 Painted metal sheets 100 1345 6 Bare soil 100 5029 7 Bitumen 100 1330 8 Self-blocking bricks 100 3682 9 Shadows 100 947
Table 2. Categories and settings of KSC dataset
View tableTable 2. Categories and settings of KSC dataset
Class Class Name Training number Total number Total 459 5211 1 Scrub 33 761 2 Willow swamp 23 243 3 CP hammock 24 256 4 CP/Oak 24 252 5 Slash pine 15 161 6 Oak/Broadleaf 22 229 7 Hardwood swamp 9 105 8 Graminoid marsh 38 431 9 Spartina marsh 51 520 10 Catiail marsh 39 404 11 Salt marsh 41 419 12 Mud flats 49 503 13 Water 91 927
Table 3. Categories and settings of Indian Pines dataset
View tableTable 3. Categories and settings of Indian Pines dataset
Class Class Name Training number Total number Total 1342 10249 1 Alfalfa 33 46 2 Corn-notill 100 1428 3 Corn-mintill 100 830 4 Corn 100 237 5 Grass-pasture 100 483 6 Grass-trees 100 730 7 Grass-pasture-mowed 20 28 8 Hay-windrowed 100 478 9 Oats 14 20 10 Soybean-notill 100 972 11 Soybean-mintill 100 2455 12 Soybean-clean 100 593 13 Wheat 100 205 14 Woods 100 1265 15 Buildings-Grass-Trees-Drives 100 386 16 Stone-Steel-Towers 75 93
Table 4. Settings of proposed method
View tableTable 4. Settings of proposed method
Method Dataset Split Hidden size Kernel size SeAMN Indian Pines 8-4-2-1 75 — Pavia University 8-4-2-1 75 — KSC 16-8-4-2 75 — SaAMN Indian Pines — — 3 Pavia University — — 3 KSC — — 3 Proposed
method
Indian Pines 8-4-2-1 75 3 Pavia University 8-4-2-1 75 3 KSC 32-16-8-4 75 3
Table 5. Classification results of different feature extraction methods on KSC dataset
View tableTable 5. Classification results of different feature extraction methods on KSC dataset
Method Spectral feature extraction method Spatial feature extraction method Joint feature extraction method 1DCNN LSTM
-byb
LSTM
-split
SeMN SeAMN 2DCNN SaMN SaAMN SSUN ASSMN Proposed
method
OA /% AP
SD
85.81
±1.26
68.11
±1.90
73.00
±1.64
88.03
±1.06
88.55
±0.82
96.58
±0.51
96.80
±1.54
97.47
±1.65
97.43
±0.65
98.01
±0.77
98.46
±0.50
AA /% AP
SD
78.90
±2.49
55.56
±3.12
61.50
±1.89
82.71
±1.36
83.23
±1.40
95.85
±0.73
95.91
±1.98
96.94
±2.21
96.84
±0.61
97.32
±1.15
98.00
±0.80
KC×100% AP
SD
84.21
±1.40
64.40
±2.18
69.93
±1.83
86.68
±1.18
87.25
±0.91
96.19
±0.56
96.44
±1.71
97.18
±1.84
97.14
±0.72
97.78
±0.86
98.29
±0.56
TrnTime /s AP
SD
7.14
±0.66
46.84
±0.43
10.93
±0.22
30.02
±0.88
50.54
±1.10
13.28
±0.32
225.88
±9.46
243.11
±6.01
96.46
±4.63
246.37
±5.83
313.15
±1.20
Classification accuracy /% 1 87.43 74.77 75.87 85.96 88.13 92.12 94.16 94.11 94.74 96.35 96.91 2 87.64 69.18 70.59 87.82 84.82 94.55 93.64 96.73 96.86 96.82 98.41 3 75.65 32.89 52.16 86.68 87.54 96.64 94.35 94.22 97.76 96.85 97.97 4 48.68 20.83 39.34 61.54 64.96 88.42 89.61 91.14 90.39 91.45 94.04 5 50.82 40.27 41.23 60.07 55.34 99.04 96.30 97.12 92.74 97.05 97.26 6 49.23 14.49 36.81 56.81 55.99 88.84 95.31 96.43 93.14 93.19 94.59 7 68.02 13.13 2.81 71.98 78.96 93.23 92.50 95.52 99.48 96.56 97.60 8 84.86 32.65 40.97 84.35 86.39 98.19 92.85 96.26 96.31 98.04 98.02 9 92.28 82.43 85.61 94.56 95.03 99.04 99.72 99.87 99.85 99.30 99.62 10 94.03 69.34 77.18 96.47 94.93 98.33 99.75 100.00 99.42 99.78 99.92 11 96.64 92.35 94.42 96.03 96.80 99.87 100.00 100.00 98.60 99.81 99.74 12 90.48 80.29 82.69 93.04 93.06 97.82 98.63 98.83 99.63 100.00 99.93 13 99.89 99.68 99.89 99.99 99.98 100.00 100.00 100.00 100.00 100.00 100.00
Table 6. Classification results of different feature extraction methods on Pavia University dataset
View tableTable 6. Classification results of different feature extraction methods on Pavia University dataset
Method Spectral feature extraction method Spatial feature extraction method Joint feature extraction method 1DCNN LSTM
-byb
LSTM
-split
SeMN SeAMN 2DCNN SaMN SaAMN SSUN ASSMN Proposed
method
OA /% AP
SD
78.37
±2.29
71.28
±6.20
69.36
±1.86
82.94
±2.47
85.91
±1.14
90.26
±1.10
92.97
±2.40
93.93
±1.26
94.71
±1.03
95.95
±1.38
96.61
±1.04
AA /% AP
SD
86.75
±0.40
80.85
±3.85
81.34
±0.91
87.72
±1.07
89.49
±0.74
93.68
±0.66
95.61
±1.18
96.02
±0.73
96.95
±0.32
97.77
±0.48
97.72
±0.51
KC×100% AP
SD
72.56
±2.54
64.36
±6.93
61.96
±1.82
77.96
±3.06
81.69
±1.45
87.27
±1.40
90.80
±3.08
92.02
±1.62
93.04
±1.33
94.66
±1.80
95.51
±1.35
TrnTime /s AP
SD
12.89
±0.73
49.43
±0.42
21.52
±0.50
61.85
±0.28
78.80
±0.59
25.90
±0.32
387.25
±4.32
416.77
±15.24
178.83
±11.70
442.47
±23.50
569.15
±2.32
Classification accuracy /% 1 77.47 70.28 71.61 82.34 82.76 88.31 90.90 92.65 93.66 95.95 95.32 2 70.94 62.28 57.89 79.94 84.22 87.47 90.44 92.10 92.41 94.15 96.10 3 82.03 59.69 70.73 79.34 81.60 88.29 89.47 90.50 95.57 97.19 95.85 4 94.43 93.44 95.76 93.49 94.53 94.84 97.20 97.82 99.42 99.17 99.16 5 99.66 97.87 98.80 99.15 99.61 99.94 99.97 99.98 99.82 99.99 99.98 6 79.84 78.60 70.81 81.88 87.19 90.34 97.09 96.25 96.56 96.74 96.52 7 92.71 90.37 91.59 91.93 91.74 97.27 99.11 99.01 97.76 99.48 99.45 8 83.67 75.24 74.94 81.52 83.90 97.57 96.43 95.98 97.59 97.29 97.08 9 100.00 99.85 99.95 99.89 99.83 99.10 99.87 99.89 99.80 100.00 99.99
Table 7. Classification results of different feature extraction methods on Indian Pines dataset
View tableTable 7. Classification results of different feature extraction methods on Indian Pines dataset
Method Spectral feature extraction method Spatial feature extraction method Joint feature extraction method 1DCNN LSTM
-byb
LSTM
-split
SeMN SeAMN 2DCNN SaMN SaAMN SSUN ASSMN Proposed
method
OA /% AP
SD
80.46
±1.41
64.26
±4.74
71.49
±1.54
75.28
±2.70
77.78
±2.40
95.87
±0.72
97.36
±0.72
97.71
±0.38
96.28
±0.78
96.79
±0.72
98.13
±0.37
AA /% AP
SD
87.44
±1.84
72.31
±4.99
79.06
±2.07
81.97
±3.07
84.16
±2.69
98.29
±0.28
98.69
±0.40
98.82
±0.25
98.50
±0.01
98.60
±0.48
99.03
±0.29
KC×100% AP
SD
77.66
±1.57
59.51
±5.22
67.58
±1.70
71.80
±3.03
74.64
±2.68
95.22
±0.83
96.94
±0.83
97.34
±0.44
95.69
±0.90
96.28
±0.83
97.84
±0.43
TrnTime /s AP
SD
20.34
±0.68
203.95
±0.66
32.92
±0.63
79.25
±0.59
110.56
±0.40
38.20
±0.33
667.89
±3.75
639.32
±38.31
273.34
±4.69
626.90
±4.94
745.69
±47.65
Classification accuracy /% 1 95.38 81.54 86.15 85.38 82.31 100.00 100.00 99.23 100.00 100.00 100.00 2 78.52 57.06 63.31 69.72 70.45 94.10 93.35 95.38 93.52 93.35 96.31 3 77.07 56.08 70.26 70.04 72.15 97.51 98.21 97.90 98.85 98.05 98.85 4 90.58 65.55 84.01 87.23 84.89 100.00 100.00 100.00 99.85 99.78 99.64 5 82.92 84.93 78.28 85.33 91.70 99.03 98.36 98.96 99.16 98.72 98.04 6 93.32 80.21 89.10 88.51 95.35 99.78 99.13 99.08 99.65 99.86 99.86 7 95.00 78.75 88.75 86.25 81.25 100.00 100.00 100.00 100.00 98.75 98.75 8 97.99 92.78 96.08 95.87 97.25 100.00 99.97 100.00 100.00 100.00 100.00 9 88.33 58.33 70.00 75.00 85.00 100.00 100.00 100.00 100.00 100.00 100.00 10 86.16 68.30 83.13 80.36 83.57 94.66 96.11 97.32 95.03 97.21 97.80 11 70.53 48.15 62.65 64.46 65.53 91.77 97.24 97.01 93.12 94.37 97.05 12 87.61 60.12 69.94 77.63 79.90 97.71 98.01 97.63 99.27 98.88 98.80 13 98.76 96.38 97.05 97.90 98.48 100.00 100.00 99.90 99.43 99.81 99.81 14 81.18 82.24 70.23 83.16 86.31 98.76 98.94 99.08 98.11 98.75 99.56 15 78.50 57.06 59.41 65.87 75.21 99.97 99.76 99.58 99.97 100.00 99.97 16 97.22 89.44 96.67 98.89 97.22 99.44 100.00 100.00 100.00 100.00 100.00
Tools
Get Citation
Copy Citation Text
Zhangchi Xu, Baofeng Guo, Wenhao Wu, Jingyun You, Xiaotong Su. Multi-Scale Feature Extraction Method of Hyperspectral Image with Attention Mechanism[J]. Laser & Optoelectronics Progress, 2024, 61(4): 0437010
Paper Information
Category: Digital Image Processing
Received: Mar. 28, 2023
Accepted: May. 4, 2023
Published Online: Feb. 26, 2024
The Author Email: Baofeng Guo (gbf@hdu.edu.cn)
CSTR:32186.14.LOP230974
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20