Hyperspectral Anomaly Detection Based on 3D Convolutional Autoencoder Network

Sheng-ming WANG; Tao WANG; Sheng-jin TANG; Yan-zhao SU

doi:10.3964/j.issn.1000-0593(2022)04-1270-08

Journals >Spectroscopy and Spectral Analysis >Volume 42 >Issue 4 >Page 1270 > Article

Spectroscopy and Spectral Analysis
Vol. 42, Issue 4, 1270 (2022)

Hyperspectral Anomaly Detection Based on 3D Convolutional Autoencoder Network

Sheng-ming WANG^1、*, Tao WANG^{1、1; *;}, Sheng-jin TANG^2、2;, and Yan-zhao SU^1、1;

Author Affiliations

¹1. Combat Support Academy, Rocket Force University of Engineering, Xi’an 710025, China

²2. Missile Engineering Academy, Rocket Force University of Engineering, Xi’an 710025, China

show less

DOI: 10.3964/j.issn.1000-0593(2022)04-1270-08 Cite this Article

Sheng-ming WANG, Tao WANG, Sheng-jin TANG, Yan-zhao SU. Hyperspectral Anomaly Detection Based on 3D Convolutional Autoencoder Network[J]. Spectroscopy and Spectral Analysis, 2022, 42(4): 1270 Copy Citation Text

show less

Fig. 1. Comparing 2D convolution operation (a) and 3D convolution operation (b)

Download full size | View in the Article

Fig. 2. 3D convolution

Download full size | View in the Article

Fig. 3. Unsupervised anomaly detection framework based on 3D-CAE

Download full size | View in the Article

Fig. 4. Results of San Diego datasets anomaly detection
(a): False color image; (b): Reference map; (c): RX; (d): SRX; (e): CRD; (f): UNRS; (g): LRASR; (h): 3D-CAEAD

Download full size | View in the Article

Fig. 5. Results of Los Angeles datasets anomaly detection
(a): False color image; (b): Reference map; (c): RX; (d): SRX; (e): CRD; (f): UNRS; (g): LRASR; (h): 3D-CAEAD

Download full size | View in the Article

Fig. 6. Results of Pavia datasets anomaly detection
(a): False color image; (b): Reference map; (c): RX; (d): SRX; (e): CRD; (f): UNRS; (g): LRASR; (h): 3D-CAEAD

Download full size | View in the Article

Fig. 7. ROC curves of San Diego datasets

Download full size | View in the Article

Fig. 8. ROC curves of Los Angeles datasets

Download full size | View in the Article

Fig. 9. ROC curves of Pavia datasets

Download full size | View in the Article

名称	传感器	波长/μm	像素	波段
San Diego-1	AVRIS	0.37~2.51	100×100	205
San Diego-2	AVRIS	0.37~2.51	100×100	191
San Diego-3	AVRIS	0.37~2.51	120×120	189
San Diego-4	AVRIS	0.37~2.51	100×100	189
Los Angeles-1	AVRIS	0.45~1.35	100×100	204
Los Angeles-2	AVRIS	0.45~1.35	100×100	207
Pavia-1	ROSIS	0.43~0.86	100×100	188
Pavia-2	ROSIS	0.43~0.86	100×100	188
Pavia-3	ROSIS	0.43~0.86	100×100	102

Table 1. Information of datasets

View in the Article

卷积层	INPUT SIZE	KERNEL SIZE	KERNEL NUMBERS	STRIDE	PADDING	OUTPUT SIZE
CONV1	224 5 5 1	2×3×3	32	2 1 1	valid	112 3 3 32
CONV2	112 3 3 32	2×1×1	64	2 1 1	valid	56 3 3 64
CONV3	56 3 3 64	2×3×3	128	2 1 1	valid	28 1 1 128
DECONV1	28 1 1 128	2×3×3	64	2 1 1	valid	56 3 3 64
DECONV2	56 3 3 64	2×1×1	32	2 1 1	valid	112 3 3 32
DECONV3	112 3 3 32	2×3×3	1	2 1 1	valid	224 5 5 1

Table 2. 3D-CAE parameter settings

View in the Article

数据集	RX	SRX	CRD	UNRS	LRASR	3D-CAEAD
San Diego-1	0.840 4	0.902 5	0.884 0	0.898 7	0.866 0	0.971 7
San Diego-2	0.952 6	0.944 1	0.768 1	0.791 7	0.962 2	0.986 9
San Diego-3	0.911 1	0.949 5	0.760 2	0.812 6	0.988 7	0.978 4
San Diego-4	0.940 3	0.961 7	0.907 1	0.919 8	0.894 9	0.983 3
Los Angeles-1	0.990 7	0.724 1	0.965 5	0.972 2	0.954 0	0.993 0
Los Angeles-2	0.944 0	0.927 5	0.922 4	0.981 6	0.989 9	0.984 6
Pavia-1	0.980 7	0.966 4	0.992 8	0.992 9	0.918 7	0.987 8
Pavia-2	0.999 1	0.995 8	0.994 4	0.997 6	0.996 3	0.999 9
Pavia-3	0.953 8	0.877 7	0.886 8	0.959 9	0.930 5	0.972 4

Table 3. AUC values of three groups of datasets

Sheng-ming WANG, Tao WANG, Sheng-jin TANG, Yan-zhao SU. Hyperspectral Anomaly Detection Based on 3D Convolutional Autoencoder Network[J]. Spectroscopy and Spectral Analysis, 2022, 42(4): 1270

Download Citation

Tools

Save the article for my favorites

Paper Information