BYOL-based self-supervised learning for hyperspectral image classification

Xizhen HAN; Zhengang JIANG; Yuanyuan LIU; Jian ZHAO; Qiang SUN; Jianzhuo LIU

doi:10.3788/IRLA20240215

Journals >Infrared and Laser Engineering >Volume 53 >Issue 10 >Page 20240215 > Article

Infrared and Laser Engineering
Vol. 53, Issue 10, 20240215 (2024)

BYOL-based self-supervised learning for hyperspectral image classification

Xizhen HAN^1,2, Zhengang JIANG¹, Yuanyuan LIU³, Jian ZHAO⁴..., Qiang SUN³ and Jianzhuo LIU³|Show fewer author(s)

Author Affiliations

¹Changchun University of Science and Technology, Changchun 130000, China

²Suzhou East Clotho Opto-Electronic Technology Co. Ltd. Zhangjiagang 215600, China

³Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Sciences, Changchun 130033, China

⁴Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, Suzhou 215000, China

show less

DOI: 10.3788/IRLA20240215 Cite this Article

Xizhen HAN, Zhengang JIANG, Yuanyuan LIU, Jian ZHAO, Qiang SUN, Jianzhuo LIU. BYOL-based self-supervised learning for hyperspectral image classification[J]. Infrared and Laser Engineering, 2024, 53(10): 20240215 Copy Citation Text

show less

Fig. 1. BYOL architecture

Download full size | View in the Article

Fig. 2. The flow chart of BSSL method proposed in this paper

Download full size | View in the Article

Fig. 3. SSTN algorithm architecture. (a) Φ search space; (b) Θ search space; (c) “AEAE” block sequence

Download full size | View in the Article

Fig. 4. Directional region generation in vertical direction. (a) Area scanned from top to bottom; (b) Area scanned from bottom to top; (c) Merged area in two directions; (d) Scanning performance example

Download full size | View in the Article

Fig. 5. Superpixel clustering result map of Indian Pines dataset. (a) Original image； (b) Edge image； (c) Superpixel clustering image

Download full size | View in the Article

Fig. 6. Indian Pines dataset. (a) Pseudo-color image; (b) Corresponding ground object type; (c) Number of sample sets

Download full size | View in the Article

Fig. 7. University of Pavia dataset. (a) Pseudo-color image; (b) Corresponding ground object type; (c) Number of sample sets

Download full size | View in the Article

Fig. 8. Salinas dataset. (a) Pseudo-color image; (b) Corresponding ground object type; (c) Number of sample sets

Download full size | View in the Article

Fig. 9. Classification maps of different methods on Indian Pines dataset

Download full size | View in the Article

Fig. 10. Classification maps of different methods on University of Pavia dataset

Download full size | View in the Article

Fig. 11. Classification maps of different methods on Salinas dataset

Download full size | View in the Article

Fig. 12. The impact of different ratios of pretraining samples on overall accuracy (OA)

Download full size | View in the Article

Fig. 13. The effect of different number of superpixel blocks on OA on the Indian Pines dataset

Download full size | View in the Article

Class	SuperPCA	S3PCA	ContrastNet	SSCL	N2SSL	BSSL
1	93.18±0.38	100.00±0.00	82.54±3.16	85.17±2.46	92.56±1.32	93.09±0.69
2	89.66±2.65	85.94±3.16	81.45±2.60	93.75±1.57	94.34±4.47	89.36±2.44
3	90.75±2.23	92.65±5.64	89.45±2.31	90.38±1.66	87.52±7.45	94.27±0.83
4	97.20±2.06	95.41±2.30	84.34±2.15	88.84±2.44	94.28±6.25	81.66±2.30
5	96.03±2.48	91.22±2.45	82.97±1.45	87.91±1.78	98.16±2.43	91.77±0.95
6	94.13±4.45	93.19±3.22	93.56±1.56	92.09±1.38	98.75±1.26	98.09±0.65
7	92.34±3.36	91.00±1.83	92.67±1.43	91.83±1.49	96.38±2.61	98.22±0.65
8	98.24±2.62	95.65±3.55	96.54±0.85	97.89±0.45	100.00±0.00	98.93±0.41
9	91.03±4.43	96.20±3.10	99.65±0.93	96.94±0.64	65.25±1.83	97.89±0.97
10	89.15±4.35	91.10±4.50	87.54±2.68	85.97±1.55	89.93±8.40	94.87±0.85
11	94.78±1.44	96.68±1.02	87.30±2.32	92.85±2.10	94.53±2.75	97.52±0.78
12	93.53±1.57	95.01±1.54	93.79±1.46	96.09±0.89	96.28±0.93	97.96±1.30
13	98.02±1.72	99.43±0.05	98.56±0.91	95.97±0.91	99.57±2.15	97.06±0.80
14	99.86±0.03	99.86±0.03	96.65±0.47	98.98±0.47	99.07±0.48	98.08±0.47
15	98.37±0.28	97.31±1.86	85.56±2.53	95.12±1.20	96.46±5.36	97.29±0.89
16	97.60±1.45	98.41±1.00	92.57±1.46	96.96±1.17	99.18±1.14	95.54±0.89
OA	(94.53±0.68)%	(94.80±1.22)%	(90.17±0.98)%	(92.73±1.07)%	(94.58±1.96)%	(95.85±0.69)%
AA	(94.62±0.74)%	(94.94±0.64)%	(90.32±1.17)%	(92.92±0.88)%	(93.89±2.25)%	(95.10±0.63)%
Kappa×100	93.73±0.78	94.96±0.74	89.36±1.21	93.12±0.83	94.27±1.67	95.02±0.76
recall	94.66±0.32	94.52±0.83	90.85±0.37	93.41±1.58	94.78±2.74	95.75±1.54
f1-score	94.65±0.74	94.02±0.55	90.38±0.89	92.45±1.35	95.38±1.85	95.67±2.85

Table 1. Classification results of different methods on Indian Pines dataset

View in the Article

Class	SuperPCA	S3PCA	ContrastNet	SSCL	N2SSL	BSSL
1	79.65±3.14	89.53±4.10	94.72±1.44	93.04±1.60	95.24±1.74	97.32±0.89
2	94.09±1.97	93.72±2.82	95.13±1.57	95.85±1.36	97.61±1.19	98.74±0.47
3	97.54±0.34	99.66±0.10	89.05±3.55	93.66±2.35	95.67±2.47	92.08±1.28
4	86.60±2.33	91.73±2.31	91.56±1.56	92.08±2.09	97.71±0.89	92.23±1.14
5	96.43±2.08	99.60±0.26	94.78±1.44	98.37±0.95	99.08±0.46	99.55±0.09
6	94.04±2.17	98.88±0.73	96.58±0.89	95.53±1.60	96.83±1.33	97.21±0.35
7	94.01±1.87	98.39±0.86	96.49±0.84	97.59±0.73	95.28±2.23	95.16±0.71
8	92.08±3.16	97.36±0.91	95.97±1.00	93.60±1.10	92.96±3.91	98.34±0.43
9	98.69±1.62	93.76±3.60	93.27±1.88	95.81±1.47	94.63±2.33	95.02±1.08
OA	(91.47±0.65)%	(96.09±1.28)%	(95.39±1.17)%	(95.62±0.90)%	(95.29±2.45)%	(95.93±0.61)%
AA	(92.57±0.45)%	(95.85±0.81)%	(94.17±1.33)%	(95.06±0.67)%	(96.11±1.84)%	(96.18±0.63)%
Kappa×100	88.80±0.82	94.86±1.66	95.7±1.45	95.41±0.64	95.34±2.31	95.91±0.47
recall	92.08±1.25	95.49±1.28	94.56±1.28	95.27±0.58	95.36±0.67	95.68±0.81
f1-score	91.20±0.84	96.26±0.56	94.82±1.92	95.34±0.72	95.44±1.64	95.82±1.26

Table 2. Classification results of different methods on University of Pavia dataset

View in the Article

Class	SuperPCA	S3PCA	ContrastNet	SSCL	N2SSL	BSSL
1	100.00±0.00	100.00±0.00	100.00±0.00	100.00±0.00	100.00±0.00	100.00±0.00
2	99.82±0.11	99.84±0.14	99.21±0.58	98.18±0.50	99.57±0.85	98.68±0.29
3	96.94±2.66	97.44±1.59	97.46±0.93	100.00±0.00	97.41±0.42	100.00±0.00
4	98.97±0.36	98.91±0.49	98.25±0.85	98.48±0.55	98.64±0.67	99.64±0.19
5	99.46±0.04	98.99±0.20	98.45±0.55	98.95±0.90	99.48±0.51	97.74±0.93
6	99.59±0.07	99.90±0.01	99.34±0.07	100.00±0.00	100.00±0.00	100.00±0.00
7	98.95±1.20	98.82±1.22	96.54±0.25	98.58±0.73	100.00±0.00	99.05±0.43
8	99.36±0.18	99.88±0.23	97.98±1.05	96.35±1.35	95.59±2.41	99.79±0.16
9	99.66±0.27	99.09±1.49	98.15±1.16	100.00±0.00	100.00±0.00	98.86±0.83
10	97.28±0.98	95.81±2.66	94.12±1.66	96.08±0.75	99.34±0.84	99.02±0.31
11	98.17±1.30	98.54±1.20	93.79±1.68	96.07±0.53	94.84±1.65	99.39±0.45
12	99.79±0.22	99.95±0.10	98.38±1.46	98.23±0.52	99.56±0.28	99.95±0.28
13	98.19±0.00	98.19±0.00	96.87±1.54	95.33±0.87	98.84±0.64	100.00±0.00
14	97.65±1.00	98.04±0.91	95.88±1.47	97.59±0.81	99.57±0.48	100.00±0.00
15	99.43±0.36	99.56±0.36	99.65±0.08	100.00±0.00	92.86±0.85	99.97±0.00
16	99.23±0.65	98.92±0.52	97.11±1.46	100.00±0.00	99.28±0.34	100.00±0.00
OA	(99.19±0.16)%	(99.16±0.49)%	(97.61±0.78)%	(98.57±0.68)%	(98.97±0.71)%	(99.52±0.27)%
AA	(98.91±0.27)%	(98.87±0.49)%	(97.57±0.52)%	(98.37±0.43)%	(98.44±0.83)%	(99.51±0.23)%
Kappa×100	99.09±0.17	99.06±0.55	97.13±0.46	98.06±0.33	97.84±0.58	99.44±0.12
recall	99.37±0.82	99.14±0.36	97.28±0.68	98.37±1.07	98.37±0.63	99.27±0.34
f1-score	98.85±0.67	98.68±0.77	97.61±1.85	98.75±0.67	97.86±1.22	99.16±0.85

Table 3. Classification results of different methods on Salinas dataset

View in the Article

Datasets	Evaluation indicators	BSSL-noSP	BSSL
Indian Pines	OA	(92.13±0.28)%	(95.85±0.69)%
	AA	(91.42±0.83)%	(95.10±0.63)%
	Kappa	91.97±0.66	95.02±0.76
University of Pavia	OA	(92.71±0.38)%	(95.93±0.61)%
	AA	(92.39±0.74)%	(96.18±0.63)%
	Kappa	92.06±0.86	95.91±0.47
Salinas	OA	(94.79±0.56)%	(99.52±0.27)%
	AA	(94.48±0.73)%	(99.51±0.23)%
	Kappa	94.27±0.22	99.44±0.12

Table 4. Ablation trial results on three datasets

Download Citation

Save the article for my favorites

Paper Information

微信扫一扫：分享

微信扫一扫：分享