Hyperspectral Image Classification Based on Edge-Preserving Filter and Deep Residual Network

Lü Huanhuan; Zhuolu Wang; Hui Zhang

doi:10.3788/LOP202259.1611005

Laser & Optoelectronics Progress, Volume. 59, Issue 16, 1611005(2022)

Hyperspectral Image Classification Based on Edge-Preserving Filter and Deep Residual Network

Lü Huanhuan^1,2, Zhuolu Wang¹, and Hui Zhang^2、*

Author Affiliations

¹School of Software, Liaoning Technical University, Huludao 125105, Liaoning , China

²School of Information Engineering, Huzhou University, Huzhou 313000, Zhejiang , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(18)

Fig. 1. Overall flow of the proposed method

Download full size

Fig. 2. Structure diagram of residual element

Download full size

Fig. 3. Model structure of depth residual network

Download full size

Fig. 4. Indian Pines dataset. (a) False color image;(b) real ground data

Download full size

Fig. 5. Pavia University dataset. (a) False color image;(b) real ground data

Download full size

Fig. 6. Classification accuracy of different dropout values. (a) Indian Pines; (b) Pavia University

Download full size

Fig. 7. Loss function and overall classification accuracy of different epoch values. (a) Indian Pines; (b) Pavia University

Download full size

Fig. 8. Classification accuracy of different $η$ values. (a) Indian Pines; (b) Pavia University

Download full size

Fig. 9. Overall classification accuracy of different $δ_{c}$ values and $δ_{s}$ values. (a) Indian Pines; (b) Pavia University

Download full size

Fig. 10. Classification results of different algorithms in Indian Pines dataset

Download full size

Fig. 11. Partial enlargement comparison of classification results of Indian Pines dataset

Download full size

Fig. 12. Classification results of different algorithms in Pavia University dataset

Download full size

Fig. 13. Partial enlargement comparison of classification results of Pavia University dataset

Download full size

Table 1. Feature map size and parameter quantities of depth residual network

View table

Table 1. Feature map size and parameter quantities of depth residual network

Input layer	Layer	Feature map size	Params
Total number of parameters		1276592
	Input	9，9，3	0
Input	C1	7，7，16	448
C1	C2	7，7，32	4640
C2	C3	7，7，32	9248
C1	R1	7，7，32	4640
C3、R1	Add1	7，7，32	0
Add1	Activation1	7，7，32	0
Activation1	C4	7，7，64	18496
C4	C5	7，7，64	36928
C1	R2	7，7，64	9280
Activation1	R4	7，7，64	18496
C5、R2、R4	Add2	7，7，64	0
Add2	Activation2	7，7，64	0
Activation2	C6	7，7，128	73856
C6	C7	7，7，128	147584
C1	R3	7，7，128	18560
Add1	R5	7，7，128	36992
Add2	R6	7，7，128	73856
C7、R3、R5、R6	Add3	7，7，128	0
Add3	Activation3	7，7，128	0
Activation3	P1	5，5，128	0
P1	Flatten1	3200	0
Flatten1	FC	256	819456
FC	Dropout	256	0
Dropout	Softmax	16	4112

Table 2. Classification accuracy corresponding to different numbers of convolution kernels
View table
Table 2. Classification accuracy corresponding to different numbers of convolution kernels
Number of convolution kernels Indian Pines Pavia University
OA /% Kappa /% OA /% Kappa /%
8 88.54 86.94 90.34 88.92
16 92.16 91.06 92.53 91.57
32 91.29 90.08 93.53 92.98
64 90.75 89.62 91.01 89.92

Table 3. Classification results of different algorithms in the Indian Pines dataset

View table

Table 3. Classification results of different algorithms in the Indian Pines dataset

Class	SP-SVM	2DCNN	Res-3DCNN	S2FEF-CNN	LBP-1DCNN	Res-2DCNN	JBF-2DCNN	JBF-Res- 2DCNN
Alfalfa	60.86	81.81	71.42	82.60	88.37	100.00	97.43	100.00
Corn-notill	76.43	78.35	91.25	90.98	94.88	98.40	98.75	96.87
Corn-min	72.89	84.26	89.79	92.65	95.29	95.04	96.58	97.36
Corn	57.56	64.95	82.43	89.04	89.37	91.54	98.03	100.00
Grass/pasture	90.11	90.42	92.80	91.87	92.93	98.57	95.55	99.08
Grass/trees	87.84	94.94	93.40	98.94	99.24	96.60	99.84	100.00
Grass-mowed	87.50	100.00	88.00	77.41	75.00	90.00	96.00	100.00
Hay-windrowed	93.18	96.57	94.05	99.53	100.00	100.00	100.00	100.00
Oats	50.00	87.50	90.90	50.00	52.17	43.47	87.50	90.00
Soybeans-notill	75.08	79.77	88.53	91.06	93.20	93.87	96.50	98.06
Soybeans-min	78.52	81.88	94.96	94.57	96.76	97.47	96.39	99.77
Soybeans-clean	83.40	81.97	88.69	91.49	92.91	96.24	97.70	97.55
Wheat	97.29	94.32	100.00	100.00	100.00	100.00	100.00	100.00
Woods	93.11	93.37	98.57	99.29	99.11	98.94	99.38	99.91
Bldg-grass-drives	77.94	75.95	89.94	93.87	94.66	96.50	95.22	99.71
Stone-steel-towes	98.57	100.00	98.61	96.92	95.52	100.00	100.00	100.00
OA	81.56	84.71	92.78	94.17	95.83	96.92	97.62	98.87
AA	80.02	86.63	90.83	90.01	91.21	93.54	97.18	98.64
Kappa	78.90	82.53	91.77	93.35	95.25	96.60	97.28	98.71

Table 4. Classification results of different algorithms in Pavia University dataset

View table

Table 4. Classification results of different algorithms in Pavia University dataset

Class	SP-SVM	2DCNN	Res-3DCNN	S2FEF-CNN	LBP-1DCNN	Res-2DCNN	JBF-2DCNN	JBF-Res-2DCNN
Asphalt	81.44	89.37	93.31	95.09	96.08	97.95	98.20	99.38
Meadows	90.26	91.67	95.12	97.83	99.23	98.36	99.78	99.92
Gravel	82.63	83.31	80.16	93.27	94.96	91.80	92.18	96.32
Trees	95.65	96.74	99.91	99.24	97.74	98.18	96.78	99.45
Painted metalsheets	99.15	99.15	91.76	99.23	100.00	97.09	100.00	98.02
Bare soil	94.11	94.23	95.08	97.74	97.41	95.74	99.59	100.00
Bitumen	94.77	87.69	83.94	83.96	85.22	98.73	95.30	98.84
Self-blocking bricks	79.26	80.90	86.07	86.52	88.15	90.23	93.54	97.67
Shadows	100.00	100.00	100.00	100.00	100.00	100.00	100.00	100.00
OA	88.58	90.86	93.28	95.85	96.78	97.14	98.26	99.35
AA	90.81	91.45	91.71	94.76	95.42	95.96	97.26	98.84
Kappa	84.53	87.69	91.01	94.48	95.73	96.48	97.69	99.13

Table 5. Training time and test time of different algorithms

View table

Table 5. Training time and test time of different algorithms

Dataset	Parameter	Algorithm
Dataset	Parameter	SP-SVM	Res-3DCNN	S2FEF-CNN	LBP-1DCNN	JBF-Res-2DCNN
Indian Pines	Training time	18.64	1507.28	1430.21	890.49	1293.52
Indian Pines	Test time	0.75	5.06	3.91	2.78	3.04
Pavia University	Training time	10.31	1002.79	921.52	629.46	862.95
Pavia University	Test time	1.42	8.35	7.28	5.05	5.93

Tools

Get Citation

Copy Citation Text

Lü Huanhuan, Zhuolu Wang, Hui Zhang. Hyperspectral Image Classification Based on Edge-Preserving Filter and Deep Residual Network[J]. Laser & Optoelectronics Progress, 2022, 59(16): 1611005

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Imaging Systems

Received: Jul. 18, 2021

Accepted: Aug. 23, 2021

Published Online: Jul. 22, 2022

The Author Email: Zhang Hui (wangzl2019@126.com)

DOI:10.3788/LOP202259.1611005

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Feature map size and parameter quantities of depth residual network

Table 1. Feature map size and parameter quantities of depth residual network

Table 2. Classification accuracy corresponding to different numbers of convolution kernels

Table 2. Classification accuracy corresponding to different numbers of convolution kernels

Table 3. Classification results of different algorithms in the Indian Pines dataset

Table 3. Classification results of different algorithms in the Indian Pines dataset

Table 4. Classification results of different algorithms in Pavia University dataset

Table 4. Classification results of different algorithms in Pavia University dataset

Table 5. Training time and test time of different algorithms

Table 5. Training time and test time of different algorithms