Semi-Supervised Hyperspectral Anomaly Detection Based on Spatial-Spectral Background Reconstruction

Luyao Li; Zhongwei Li; Leiquan Wang; Juan Li; Shunxiao Shi

doi:10.3788/LOP222705

Laser & Optoelectronics Progress, Volume. 60, Issue 20, 2028001(2023)

Semi-Supervised Hyperspectral Anomaly Detection Based on Spatial-Spectral Background Reconstruction

Luyao Li¹, Zhongwei Li^1、*, Leiquan Wang², Juan Li², and Shunxiao Shi²

Author Affiliations

¹College of Oceanography and Space Informatics, China University of Petroleum, Qingdao 266580, Shandong , China

²College of Computer Science and Technology, China University of Petroleum, Qingdao 266580, Shandong , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(16)

Fig. 1. Hyperspectral anomaly detection model combined with spatial-spectrum

Download full size

Fig. 2. Flow chart of cluster generation

Download full size

Fig. 3. Loss architecture diagrams. (a) Potential feature layer constraint loss; (b) reconstruction loss

Download full size

Fig. 4. Loss architecture

Download full size

Fig. 5. Filtering results in spatial domain

Download full size

Fig. 6. Visual inspection maps of different methods. (a) Sandiego; (b) ABU; (c) TC-1; (d) TC-2; (e) BC

Download full size

Fig. 7. ROC curves of different methods. (a) Sandiego; (b) ABU; (c) TC-1; (d) TC-2; (e) BC

Download full size

Fig. 8. Background anomaly separation diagram. (a) Sandiego; (b) ABU; (c) TC-1; (d) TC-2; (e) BC

Download full size

Fig. 9. Spectral reconstruction at different positions. (a) Different position distribution; (b) original spectral diagram; (c) reconstructed spectral map

Download full size

Fig. 10. Visual effect images reconstructed from different batches. (a) Spectral; (b) spatial

Download full size

Fig. 11. Influence of the number of clusters c on different datasets

Download full size

Fig. 12. Effects of equilibrium coefficients α and β on different datasets. (a) Sandiego; (b) ABU; (c) TC-1; (d) TC-2; (e) BC

Download full size

Table 1. AUC values on different datasets

View table

Table 1. AUC values on different datasets

Method	Sandiego	ABU	TC-1	TC-2	BC
RX	0.9600	0.8403	0.9906	0.9945	0.9997
LRX	0.8730	0.9534	0.9661	0.9402	0.9999
LRASR	0.9933	0.8857	0.9850	0.9956	0.9987
LSMAD	0.9756	0.9184	0.9803	0.9852	0.9996
GTVLRR	0.9749	0.8958	0.9773	0.9809	0.9937
RPCA	0.9673	0.8431	0.9922	0.9957	0.9995
PTA	0.9877	0.9176	0.9720	0.9941	0.9823
Proposed method	0.9975	0.9909	0.9983	0.9913	0.9996

Table 2. Parameter settings on different datasets

View table

Table 2. Parameter settings on different datasets

Method

Sandiego

ABU

TC-1

TC-2

LRX

$w_{o u t} = 25$

$W_{i n e r} = 23$

$w_{o u t} = 25$

$W_{i n e r} = 23$

$w_{o u t} = 15$

$W_{i n e r} = 13$

$w_{o u t} = 25$

$W_{i n e r} = 23$

$w_{o u t} = 25$

$W_{i n e r} = 23$

LRASR

$λ = 0.01$

$β = 0.005$

$λ = 0.2$

$β = 0.1$

$λ = 0.01$

$β = 0.005$

$λ = 0.001$

$β = 1$

$λ = 1$

$β = 0.1$

GTVLRR

$λ = 0.01$

$β = 0.005$

$γ = 0.05$

$λ = 0.5$

$β = 0.2$

$γ = 0.02$

$λ = 0.01$

$β = 0.005$

$γ = 0.05$

$λ = 0.005$

$β = 0.2$

$γ = 0.05$

$λ = 0.5$

$β = 0.2$

$γ = 0.05$

PTA

rank 1

$α = 2$

$β = 0.05$

$τ = 0.0001$

$μ = 0.001$

rank 1

$α = 2$

$β = 0.05$

$τ = 0.1$

$μ = 0.001$

rank 1

$α = 2$

$β = 0.0001$

$τ = 0.1$

$μ = 0.001$

rank 1

$α = 2$

$β = 0.05$

$τ = 0.1$

$μ = 0.001$

rank 1

$α = 2$

$β = 0.05$

$τ = 0.0001$

$μ = 0.001$

Proposed method

c=10

$θ =$ 0.003

$α = 1$

$β = 0.9$

c=15

$θ =$ 0.005

$α = 1$

$β = 0.9$

c=10

$θ =$ 0.003

$α = 0.9$

$β = 0.8$

c=10

$θ =$ 0.005

$α = 0.9$

$β = 0.9$

c=10

$θ =$ 0.005

$α = 0.9$

$β = 0.8$

Table 3. AUC values of the eight versions

View table

Table 3. AUC values of the eight versions

Method	Sandiego	ABU	TC-1	TC-2	BC
Spectral GAN + Filter	0.9686	0.9601	0.9669	0.9705	0.9753
Spectral GAN + Spatial GAN	0.9779	0.9615	0.9814	0.9843	0.9854
Spectral GAN + Spatial GAN + Filter	0.9815	0.9788	0.9898	0.9764	0.9805
FCM-1 + Spectral GAN + Filter	0.9696	0.9736	0.9812	0.9814	0.9827
FCM-1 + Spectral GAN + Spatial GAN	0.9735	0.9666	0.9679	0.9766	0.9864
FCM-1 + Spectral GAN + Spatial GAN + Filter	0.9789	0.9789	0.9900	0.9770	0.9931
FCM-1 + Spectral GAN+FCM-2 + Spatial GAN	0.9911	0.9657	0.9836	0.9882	0.9899
FCM-1 + Spectral GAN+FCM-2+ Spatial GAN+ Filter	0.9975	0.9909	0.9983	0.9913	0.9996

Table 4. Time costs of different methods

View table

Table 4. Time costs of different methods

Method	Sandiego	ABU	TC-1	TC-2	BC
RX	0.0301	0.0899	0.0869	0.0946	0.0718
LRX	13.3475	41.5584	32.8602	41.3030	32.5046
LRASR	20.8709	39.1473	41.6164	45.9174	43.5895
LSMAD	5.2790	11.0777	11.8113	11.3911	10.3939
GTVLRR	274.4787	296.2252	191.6086	136.9443	317.7260
RPCA	7.0242	15.3004	7.1714	7.7943	5.5224
PTA	15.4195	35.5699	30.4078	32.2381	29.9787
Proposed method	7.3677	7.5819	6.1355	6.1538	6.3809

Tools

Get Citation

Copy Citation Text

Luyao Li, Zhongwei Li, Leiquan Wang, Juan Li, Shunxiao Shi. Semi-Supervised Hyperspectral Anomaly Detection Based on Spatial-Spectral Background Reconstruction[J]. Laser & Optoelectronics Progress, 2023, 60(20): 2028001

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. 8, 2022

Accepted: Nov. 29, 2022

Published Online: Oct. 13, 2023

The Author Email: Zhongwei Li (li.zhongwei@vip.163.com)

DOI:10.3788/LOP222705

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology