Dual-Stream Convolutional Autoencoding Network for Hyperspectral Unmixing using Attention Mechanism

Xiaotong Su; Baofeng Guo; Jingyun You; Wenhao Wu; Zhangchi Xu

doi:10.3788/LOP231022

Laser & Optoelectronics Progress, Volume. 61, Issue 4, 0428008(2024)

Dual-Stream Convolutional Autoencoding Network for Hyperspectral Unmixing using Attention Mechanism

Xiaotong Su, Baofeng Guo^*, Jingyun You, Wenhao Wu, and Zhangchi Xu

Author Affiliations

School of Automation, Hangzhou Dianzi University, Hangzhou 310018, Zhejiang, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(11)

Fig. 1. Network architecture of proposed method

Download full size

Fig. 2. Architecture of proposed channel attention mechanism

Download full size

Fig. 3. Abundance maps of trees, water, soil, and roads on Jasper Ridge dataset obtained by different algorithms

Download full size

Fig. 4. Abundance maps of soil, trees, and water on Samson dataset obtained by different algorithms

Download full size

Fig. 5. Abundance maps corresponding to different endmembers in Cuprite dataset obtained by different algorithms and the pseudo color image of Cuprite dataset

Download full size

Table 1. Configuration of the feature fusion network

View table

Table 1. Configuration of the feature fusion network

Spectral stream convolution network	Spatial stream convolution network
Conv2D，size is 1×1，number of channels is 128，padding is 1	Conv2D，size is 3×3，number of channels is 128，padding is 1
BatchNorm2d	BatchNorm2d
Dropout	Dropout
ReLU	ReLU
Conv2D，size is 1×1，number of channels is 64，padding is 1	Conv2D，size is 1×1，number of channels is 64，padding is 1
BatchNorm2d	BatchNorm2d
	SEAttention

Table 2. Configuration of the spectral unmixing network

View table

Table 2. Configuration of the spectral unmixing network

Encoder layer	Conv2D，size is $1 \times 1$ ，number of channels is 64，padding is 0
	BatchNorm2d
	Dropout
	ReLU
	Conv2D，size is $1 \times 1$ ，number of channels is P，padding is 1
	ReLU
Decoder layer	Conv2D，size is $1 \times 1$ ，number of channels is P，padding is 1
Decoder layer	ReLU

Table 3. Parameter setting in our experiment
View table
Table 3. Parameter setting in our experiment
Dataset $α$ $β$ $λ$ $θ$ Learning rate Epoch
Samson 0.3 0.3 0 1×10^-4 0.008 520
Jasper Ridge 0.6 0.8 0.4 1×10^-7 0.008 280

Table 4. Quantitative results of different algorithms on Jasper Ridge dataset

View table

Table 4. Quantitative results of different algorithms on Jasper Ridge dataset

Parameter		$L_{1 / 2} - N M F$	TV-RSNMF	CNNAEU	CyCU-Net	AAS	MAC-U	Proposed algorithm
SAD	Tree	0.2046	0.0340	0.1420	0.0399	0.0387	0.1546	0.0985
	Water	0.2729	0.1583	0.4159	0.1527	0.1330	0.0900	0.0341
	Soil	0.0799	0.0306	0.0526	0.0306	0.1554	0.1175	0.0753
	Road	0.0691	0.0692	0.1005	0.0392	0.0456	0.0913	0.0243
Mean SAD		0.1566	0.0730	0.1778	0.0656	0.0932	0.1134	0.0580
RMSE		0.1420	0.1131	0.1790	0.1163	0.1233	0.1599	0.0811

Table 5. Quantitative results of different algorithms on Samson dataset

View table

Table 5. Quantitative results of different algorithms on Samson dataset

Parameter		$L_{1 / 2}$ - $N M F$	TV-RSNMF	CNNAEU	CyCU-Net	AAS	MAC-U	Proposed algorithm
SAD	Soil	0.0403	0.0166	0.1954	0.0102	0.0342	0.0195	0.0143
	Tree	0.0848	0.0291	0.1100	0.0250	0.0697	0.0391	0.0326
	Water	0.2884	0.2462	0.2455	0.0520	0.1760	0.1562	0.0386
Mean SAD		0.1378	0.0973	0.1836	0.0290	0.0933	0.0716	0.0285
RMSE		0.1668	0.1904	0.1531	0.1779	0.1228	0.3293	0.07404

Table 6. SAD values of different algorithms on Cuprite dataset

View table

Table 6. SAD values of different algorithms on Cuprite dataset

Sequence	MAC-U	CNNAEU	Proposed algorithm
Mean	0.1743	0.1395	0.1701
#1Alunite	0.0963	0.4665	1.0038
#2Andradite	0.0729	0.0809	0.1001
#3Buddingtonite	0.1177	0.1114	0.1584
#4Dumortierite	0.1134	0.1241	0.1004
#5Kaolinite1	0.2312	0.1744	0.0819
#6Kaolinite2	0.0966	0.0709	0.0882
#7Muscovite	0.1140	0.1768	0.1084
#8Montmorillonite	0.0874	0.0873	0.0719
#9Nontronite	0.0960	0.0789	0.0805
#10Pyrope	0.1160	0.0838	0.0588
#11Sphene	0.8368	0.0720	0.0703
#12Chalcedony	0.1142	0.1380	0.1188

Tools

Get Citation

Copy Citation Text

Xiaotong Su, Baofeng Guo, Jingyun You, Wenhao Wu, Zhangchi Xu. Dual-Stream Convolutional Autoencoding Network for Hyperspectral Unmixing using Attention Mechanism[J]. Laser & Optoelectronics Progress, 2024, 61(4): 0428008

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites