Hyperspectral Image Classification Algorithm Based on Two-Channel Generative Adversarial Network

Xiaojun Bi; Zeyu Zhou

doi:10.3788/AOS201939.1028002

Journals >Acta Optica Sinica >Volume 39 >Issue 10 >Page 1028002 > Article

Acta Optica Sinica
Vol. 39, Issue 10, 1028002 (2019)

Hyperspectral Image Classification Algorithm Based on Two-Channel Generative Adversarial Network

Xiaojun Bi¹ and Zeyu Zhou^2、*

Author Affiliations

¹ Department of Information Engineering, Minzu University of China, Beijing 100081, China

² College of Information and Communication Engineering, Harbin Engineering University, Harbin, Heilongjiang 150001, China

show less

DOI: 10.3788/AOS201939.1028002 Cite this Article Set citation alerts

Xiaojun Bi, Zeyu Zhou. Hyperspectral Image Classification Algorithm Based on Two-Channel Generative Adversarial Network[J]. Acta Optica Sinica, 2019, 39(10): 1028002 Copy Citation Text

show less

Fig. 1. GAN framework structure

Download full size

Fig. 2. Structural diagram of DCGAN on LSUN dataset

Download full size

Fig. 3. GAN classification framework of hyperspectral image

Download full size

Fig. 4. Improved one-dimensional GAN classification structure

Download full size

Fig. 5. Improved two-dimensional GAN classification structure

Download full size

Fig. 6. Two-channel GAN classification structure

Download full size

Fig. 7. Salinas dataset. (a) Pseudo color composite map; (b) feature reference map

Download full size

Fig. 8. Classification results of the eight algorithms on the Salinas dataset. (a) Real feature reference map; (b) 1D-CNN; (c) 1D-GAN; (d) HS-1D-GAN; (e) 2D-CNN; (f) HS-2D-GAN; (g) 3D-CNN; (h) 3D-GAN; (i) HS-TC-GAN

Download full size

Fig. 9. Indian pines dataset.(a) Pseudo color composite map; (b) feature reference map

Download full size

Fig. 10. Classification results of the eight algorithms on the Indian pines dataset. (a) Real feature reference map; (b) 1D-CNN; (c) 1D-GAN; (d) HS-1D-GAN; (e) 2D-CNN; (f) HS-2D-GAN; (g) 3D-CNN; (h) 3D-GAN; (i) HS-TC-GAN

Download full size

Networks	Layer	Operation	Kernel size	BN	Stride	Padding	Activation function
Generator	1	Deconv	1×1×1024	No	1	0	ReLU
	2	Deconv	1×1×128×a	Yes	1	0	ReLU
	3	Reshape	-	No	-	-	No
	4	Deconv	4×1×256	Yes	2	1	ReLU
	5	Deconv	4×1×64	Yes	2	1	ReLU
	6	Deconv	1×1×1	No	1	0	Tanh
	7	Full	1×n_nc	No	-	-	No
Discriminator	1	Conv	3×1×32	No	1	1	LeakyReLU
	2	Conv	3×1×32	No	1	1	LeakyReLU
	2	Conv	3×1×64	No	1	0	LeakyReLU
	3	Conv	3×1×64	No	1	0	LeakyReLU
	3	Conv	3×1×128	No	2	1	LeakyReLU
	4	Conv	3×1×128	No	2	1	LeakyReLU
	4	Conv	3×1×256	No	1	0	LeakyReLU
	5	Conv	3×1×256	No	1	0	LeakyReLU
	5	Conv	3×1×128	No	1	0	LeakyReLU
	6	Conv	3×1×128	No	1	0	LeakyReLU
	6	Conv	3×1×32	No	2	1	LeakyReLU
	7	Conv	3×1×32	No	2	1	LeakyReLU
	7	Reshape	-	No	-	-	No
	8	Reshape	-	No	-	-	No
	8	Conv	1×1×1024	No	1	0	No
	9	Conv	1×1×1024	No	1	0	No
		Softmax	1024×n_nclass	No	-	-	No
		Sigmoid	1024×2	No	-	-	No

Table 1. Improved one-dimensional GAN classification framework

Networks	Layer	Operation	Kernel size	BN	Stride	Padding	Activation function
Generator	1	Deconv	1×1×1024	No	1	0	ReLU
	2	Reshape	-	No	-	-	No
	3	Deconv	4×4×128	Yes	2	1	ReLU
	4	Deconv	4×4×256	Yes	2	1	ReLU
	5	Deconv	4×4×128	Yes	2	1	ReLU
	6	Deconv	4×4×3	No	2	1	Tanh
Discriminator	1	Conv	3×3×32	No	2	1	LeakyReLU
	2	Conv	3×3×64	No	2	1	LeakyReLU
	3	Conv	3×3×128	No	2	1	LeakyReLU
	4	Conv	3×3×64	No	2	1	LeakyReLU
	5	Reshape	-	No	-	-	No
	6	Conv	1×1×1024	No	1	0	No
	7	Softmax	1024×n_nclass	No	-	-	No
	7	Sigmoid	1024×2	No	-	-	No

Table 2. Improved two-dimensional GAN classification framework

No.	Class	Sample number
1	Brocoli_green_weeds_1	1977
2	Brocoli_green_weeds_2	3726
3	Fallow	1976
4	Fallow_rough_plow	1394
5	Fallow_smooth	2678
6	Stubble	3959
7	Celery	3579
8	Grapes_untrained	11213
9	Soil_vinyard_develop	6197
10	Corn_senesced_green_weeds	3249
11	Lettuce_romaine_4wk	1058
12	Lettuce_romaine_5wk	1908
13	Lettuce_romaine_6wk	909
14	Lettuce_romaine_7wk	1061
15	Vinyard_untrained	7164
16	Vinyard_vertical_trellis	1737
Total		53785

Table 3. [in Chinese]

Index	1D-CNN	1D-GAN	HS-1D-GAN	2D-CNN	HS-2D-GAN	3D-CNN	3D-GAN	HS-TC-GAN
OA /%	86.12	86.88	90.22	87.84	97.15	92.04	93.38	99.67
AA /%	89.63	92.24	94.05	89.86	96.96	94.54	95.20	99.45
Kappa /%	84.48	85.39	89.10	86.44	96.82	91.13	92.63	99.63
Train time /s	8.96	19.67	120.99	94.61	195.60	211.90	350.49	385.27
Test time /s	0.51	0.54	2.87	3.95	2.43	4.00	3.48	5.11
Total time /s	9.47	20.21	123.86	98.56	198.03	215.90	353.97	390.38

Table 4. Comparison of classification performances of eight algorithms on Salinas dataset

No.	Class	Sample number
1	Alfalfa	46
2	Corn-notill	1428
3	Corn-min	830
4	Corn	237
5	Grass-pasture	483
6	Grass-trees	730
7	Grass-pasture-mowed	28
8	Hay-windrowed	478
9	Oats	20
10	Soybean-notill	972
11	Soybean-mintill	2455
12	Soybean-clean	593
13	Wheat	205
14	Woods	1265
15	Buildings-Grass-Trees	386
16	Stone-Steel-Towers	93
Total		10249

Table 5. [in Chinese]

Index	1D-CNN	1D-GAN	HS-1D-GAN	2D-CNN	HS-2D-GAN	3D-CNN	3D-GAN	HS-TC-GAN
OA /%	62.55	63.92	68.90	92.38	94.36	92.50	93.34	99.74
AA /%	53.29	55.94	60.28	84.87	91.12	89.78	86.80	97.52
Kappa /%	56.51	58.41	64.29	91.29	93.57	91.44	92.39	99.70
Train time /s	9.06	19.46	122.06	92.36	194.33	201.48	333.63	386.05
Test time /s	0.13	0.13	1.96	0.84	0.55	0.93	0.72	2.07
Total time /s	9.19	19.59	124.02	93.20	194.88	202.41	334.39	388.12

Table 6. Comparison of classification performances of eight algorithms on Indian pines dataset

Xiaojun Bi, Zeyu Zhou. Hyperspectral Image Classification Algorithm Based on Two-Channel Generative Adversarial Network[J]. Acta Optica Sinica, 2019, 39(10): 1028002

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information