Hyperspectral Images Classification Method Based on 3D Octave Convolution and Bi-RNN Attention Network

Lianhui LIANG; Jun LI; Shaoquan ZHANG

doi:10.3788/gzxb20215009.0910001

Journals >Acta Photonica Sinica >Volume 50 >Issue 9 >Page 0910001 > Article

Acta Photonica Sinica
Vol. 50, Issue 9, 0910001 (2021)

Hyperspectral Images Classification Method Based on 3D Octave Convolution and Bi-RNN Attention Network

Lianhui LIANG¹, Jun LI², and Shaoquan ZHANG^1、*

Author Affiliations

¹College of Electrical and Information Engineering， Hunan University， Changsha40082， China

²School of Geography and Planning， Sun Yat-sen University， Guangzhou51075， China

show less

DOI: 10.3788/gzxb20215009.0910001 Cite this Article

Lianhui LIANG, Jun LI, Shaoquan ZHANG. Hyperspectral Images Classification Method Based on 3D Octave Convolution and Bi-RNN Attention Network[J]. Acta Photonica Sinica, 2021, 50(9): 0910001 Copy Citation Text

show less

Fig. 1. Flow of framework for 3D Octave convolution and Bi-RNN attention network

Download full size | View in the Article

Fig. 2. Flow of 3D Octave convolution network

Download full size | View in the Article

Fig. 3. Structure of Bi-RNN network

Download full size | View in the Article

Fig. 4. Structure of Bi-RNN attention network

Download full size | View in the Article

Fig. 5. The whole structure diagram of 3D Octave convolution and Bi-RNN attention network

Download full size | View in the Article

Fig. 6. Classification maps of different methods on the Pavia University dataset

Download full size | View in the Article

Fig. 7. Partial enlargement comparison of classification maps on the Pavia University dataset

Download full size | View in the Article

Fig. 8. Classification maps of different methods on the Botswana dataset

Download full size | View in the Article

Fig. 9. Partial enlargement comparison of classification maps on the Botswana dataset

Download full size | View in the Article

Code	Class name	Train	Test	Total
Total		3 930	38 846	42 776
1	Asphalt	548	6 083	6 631
2	Meadows	540	18 109	18 649
3	Gravel	392	1 707	2 099
4	Trees	542	2 522	3 064
5	Painted metal sheets	256	1 089	1 345
6	Bare soil	532	4 497	5 029
7	Bitumen	375	955	1 330
8	Self-Blocking bricks	514	3 168	3 682
9	Shadows	231	716	947

Table 1. Number of training and testing samples of the Pavia University dataset

View in the Article

Code	Class name	Train	Test	Total
Total		420	2 828	3 248
1	Water	30	240	270
2	Hippo grass	30	71	101
3	Floodplain grasses1	30	221	251
4	Floodplain grasses1	30	185	215
5	Reeds1	30	239	269
6	Riparian	30	239	269
7	Firescar2	30	229	259
8	Island interior	30	173	203
9	Acacia woodlands	30	284	314
10	Acacia shrublands	30	218	248
11	Acacia grasslands	30	275	305
12	Short mopane	30	151	181
13	Mixed mopane	30	238	268
14	Exposed soils	30	65	95

Table 2. Number of training and testing samples of the Botswana dataset

View in the Article

	Spatial size
	11 $\times$ 11	13 $\times$ 13	15 $\times$ 15	17 $\times$ 17	19 $\times$ 19	21 $\times$ 21
Accuracy OA /%	99.86	99.93	99.97	99.96	99.96	99.91

Table 3. Classification accuracy of different spatial size

View in the Article

	Dropout
	0.2	0.3	0.4	0.5	0.6	0.7	0.8	0.9
Accuracy OA /%	99.82	99.87	99.91	99.94	99.97	99.96	99.96	99.95

Table 4. Classification accuracy of different dropout

View in the Article

Class	SVM	DBMA	ARNN	SSAN	3DOC-SSAN	3DOC-RNN
1	89.33	99.98	98.93	99.28	99.82	99.93
2	93.84	99.96	96.82	98.66	99.93	99.99
3	85.82	99.91	97.71	98.53	99.67	99.88
4	97.86	99.18	99.72	98.66	99.81	99.96
5	98.99	99.90	100.00	100.00	100.00	100.00
6	94.95	99.30	99.84	99.40	100.00	100.00
7	94.14	99.92	99.69	99.79	99.76	100.00
8	89.96	96.82	99.72	99.75	99.80	99.94
9	99.98	99.33	100.00	99.72	100.00	100.00
OA/%	93.12	99.52	98.23	99.00	99.87	99.97
AA/%	90.70	99.37	99.17	99.30	99.87	99.96
Kappa/%	93.88	99.35	97.59	98.64	99.94	99.96

Table 5. Classification performance of different methods on Pavia University dataset

View in the Article

Class	SVM	DBMA	ARNN	SSAN	3DOC-SSAN
OA/%	+6.85	+0.45	+1.74	+0.97	+0.10
AA/%	+9.26	+0.59	+0.79	+0.66	+0.09
Kappa/%	+6.08	+0.61	+1.47	+2.37	+0.02

Table 6. Comparison with other methods in the classification accuracy difference on Botswana dataset

View in the Article

Class	SVM	DBMA	ARNN	SSAN	3DOC-SSAN	3DOC-RNN
1	100.00	98.94	99.60	99.19	100.00	100.00
2	100.00	100.00	100.00	100.00	100.00	100.00
3	98.19	100.00	99.13	99.15	100.00	97.38
4	96.76	99.03	88.89	100.00	100.00	100.00
5	76.15	98.81	84.62	85.49	96.65	100.00
6	74.48	98.73	99.19	98.39	97.91	99.58
7	95.63	100.00	97.05	100.00	100.00	100.00
8	98.84	100.00	100.00	99.45	100.00	99.94
9	85.21	100.00	98.97	96.56	100.00	100.00
10	89.91	100.00	94.74	100.00	100.00	100.00
11	95.27	100.00	100.00	100.00	100.00	100.00
12	92.05	100.00	100.00	100.00	98.01	100.00
13	88.24	100.00	100.00	100.00	100.00	100.00
14	100.00	100.00	97.14	95.38	100.00	100.00
OA/%	90.91	99.62	97.02	98.00	99.43	99.79
AA/%	90.14	99.67	97.10	98.12	99.47	99.81
Kappa/%	92.20	99.59	96.76	97.83	99.39	99.77

Table 7. Classification performance of different methods

View in the Article

Class	SVM	DBMA	ARNN	SSAN	3DOC-SSAN
OA/%	+8.88	+0.17	+2.77	+1.79	+0.36
AA/%	+9.67	+0.14	+2.71	+1.69	+0.34
Kappa/%	+7.57	+0.18	+3.01	+1.94	+0.38

Table 8. Comparison with other methods in the classification accuracy difference on Botswana dataset

View in the Article

Class	3D Octave	Bi-RNN	3D-OCM	SSAM+SSICM
3DOC-RNN/s	26.212	16.062	—	—
3DOC-SSAN/s	—	—	26.635	0.766 9

Table 9. Comparison of the running time of the two methods

Lianhui LIANG, Jun LI, Shaoquan ZHANG. Hyperspectral Images Classification Method Based on 3D Octave Convolution and Bi-RNN Attention Network[J]. Acta Photonica Sinica, 2021, 50(9): 0910001

Download Citation

Tools

Save the article for my favorites

Paper Information