Lightweight Residual Network Based on Depthwise Separable Convolution for Hyperspectral Image Classification

Rongjie Cheng; Yun Yang; Longwei Li; Yanting Wang; Jiayu Wang

doi:10.3788/AOS221848

Acta Optica Sinica, Volume. 43, Issue 12, 1228010(2023)

Lightweight Residual Network Based on Depthwise Separable Convolution for Hyperspectral Image Classification

Rongjie Cheng¹, Yun Yang^1,2、*, Longwei Li¹, Yanting Wang¹, and Jiayu Wang¹

Author Affiliations

¹School of Geological Engineering and Geomatics, Chang'an University, Xi'an 710054, Shaanxi, China

²Key Laboratory of Disaster Mechanism and Prevention of Mine Geological Disasters, Ministry of Natural Resources, Xi'an 710054, Shaanxi, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(15)

Fig. 1. Structure of depthwise separable convolution

Download full size

Fig. 2. Residual structure in the model of Ref. [16]

Download full size

Fig. 3. Depthwise separable convolution dotted shortcuts structure

Download full size

Fig. 4. Depthwise separable convolution identity shortcuts structure

Download full size

Fig. 5. Proposed network structure

Download full size

Fig. 6. Classification accuracy of attenuation parameters with different weights

Download full size

Fig. 7. Classification accuracy of different batchsize

Download full size

Fig. 8. Classification accuracy of proposed model with different convolution kernel sizes

Download full size

Fig. 9. Classification results of different models on the Indian Pines dataset. (a) Ground truth; (b) 3DCNN-DSC; (c) 2D-Res-CNN;(d) 3D-Res-CNN; (e) Res14; (f) DSC-Res14

Download full size

Fig. 10. Classification results of different models on the Pavia University dataset. (a) Ground truth; (b) 3DCNN-DSC; (c) 2D-Res-CNN; (d) 3D-Res-CNN; (e) Res14; (f) DSC-Res14

Download full size

Table 1. Category No. and sample size in Indian Pines dataset

View table

Table 1. Category No. and sample size in Indian Pines dataset

No.	Category	Sample size	No.	Category	Sample size
1	Alfalfa	46	9	Oats	20
2	Corn-notill	1428	10	Soybean-notill	972
3	Corn-mintill	830	11	Soybean-mintill	2455
4	Corn	237	12	Soybean-clean	593
5	Grass-pasture	483	13	Wheat	205
6	Grass-trees	730	14	Woods	1265
7	Grass-pasture-mowed	28	15	Buildings-grass-trees-drives	386
8	Hay-windrowed	478	16	stone-steel-towers	93

Table 2. Category No. and sample size in Pavia University dataset
View table
Table 2. Category No. and sample size in Pavia University dataset
No. Category Sample size No. Category Sample size
1 Asphalt 6631 6 Bare soil 5029
2 Meadows 18649 7 Bitumen 1330
3 Gravel 2099 8 Self-blocking bricks 3682
4 Trees 3064 9 Shadows 947
5 Painted metal sheets 1345

Table 3. Comparison of Conv parameter, parameter number, FLOPs, and training time of each network

View table

Table 3. Comparison of Conv parameter, parameter number, FLOPs, and training time of each network

Model	3DCNN-DSC	2D-Res-CNN	3D-Res-CNN	Res14	DSC-Res14
Conv parameter	200228	611840	111424	152232	20482
Parameter number	200228	611840	136016	281272	149522
FLOPs /10⁶	206	2	291	347	50
Time（IP）/s	300.30	21.32	1165.10	1170.32	394.71
Time（PU）/s	1913.00	25.97	997.12	1830.24	641.63

Table 4. Comparison of classification accuracy of different algorithms on the Indian Pines test set

View table

Table 4. Comparison of classification accuracy of different algorithms on the Indian Pines test set

No.	3DCNN-DSC	2D-Res-CNN	3D-Res-CNN	Res14	DSC-Res14
1	92.31	69.81	61.61	100.00	96.92
2	91.22	95.18	97.52	99.08	99.21
3	97.41	90.58	96.49	99.56	99.36
4	97.59	89.75	98.62	99.72	99.45
5	98.41	98.85	98.58	99.53	99.23
6	98.95	99.29	97.50	99.19	99.28
7	100.00	67.78	63.36	100.00	100.00
8	95.96	96.15	99.31	100.00	100.00
9	80.00	69.17	58.33	96.03	94.67
10	98.90	94.98	97.05	99.97	99.71
11	97.86	91.67	92.37	99.33	99.50
12	89.62	96.92	91.25	99.71	99.47
13	98.80	99.10	98.16	100.00	100.00
14	98.82	97.81	98.75	99.97	99.75
15	99.33	93.74	98.99	99.91	99.49
16	89.74	94.31	53.54	98.56	98.41
OA	96.46	84.41	96.09	99.51	99.46
AA	95.31	90.32	87.59	99.41	99.03
Kappa	95.96	93.61	95.53	99.10	99.38

Table 5. Comparison of classification accuracy of different algorithms on the Pavia University test set

View table

Table 5. Comparison of classification accuracy of different algorithms on the Pavia University test set

No.	3DCNN-DSC	2D-Res-CNN	3D-Res-CNN	Res14	DSC-Res14
1	95.52	93.08	98.17	99.71	99.77
2	99.92	97.75	99.16	99.93	99.83
3	100.00	87.51	99.55	99.62	99.54
4	99.75	99.19	98.13	98.37	98.92
5	99.82	99.95	99.83	100.00	99.47
6	92.49	95.14	99.98	99.98	99.98
7	97.86	93.23	98.25	99.57	99.47
8	94.73	83.66	96.60	98.38	99.31
9	99.73	97.80	99.05	99.18	98.68
OA	97.65	94.90	98.79	99.61	99.65
AA	97.75	94.15	98.75	99.42	99.44
Kappa	96.90	93.23	98.39	99.49	99.53

Tools

Get Citation

Copy Citation Text

Rongjie Cheng, Yun Yang, Longwei Li, Yanting Wang, Jiayu Wang. Lightweight Residual Network Based on Depthwise Separable Convolution for Hyperspectral Image Classification[J]. Acta Optica Sinica, 2023, 43(12): 1228010

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Remote Sensing and Sensors

Received: Oct. 19, 2022

Accepted: Dec. 12, 2022

Published Online: Jun. 20, 2023

The Author Email: Yang Yun (yangyunbox@chd.edu.cn)

DOI:10.3788/AOS221848

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Category No. and sample size in Indian Pines dataset

Table 1. Category No. and sample size in Indian Pines dataset

Table 2. Category No. and sample size in Pavia University dataset

Table 2. Category No. and sample size in Pavia University dataset

Table 3. Comparison of Conv parameter, parameter number, FLOPs, and training time of each network

Table 3. Comparison of Conv parameter, parameter number, FLOPs, and training time of each network

Table 4. Comparison of classification accuracy of different algorithms on the Indian Pines test set

Table 4. Comparison of classification accuracy of different algorithms on the Indian Pines test set

Table 5. Comparison of classification accuracy of different algorithms on the Pavia University test set

Table 5. Comparison of classification accuracy of different algorithms on the Pavia University test set