Research on the Improvement of Spectra Classification Performance With the High-Performance Hybrid Deep Learning Network

Zhong-bao LIU; Jie WANG

doi:10.3964/j.issn.1000-0593(2022)03-0699-05

Journals >Spectroscopy and Spectral Analysis >Volume 42 >Issue 3 >Page 699 > Article

Spectroscopy and Spectral Analysis
Vol. 42, Issue 3, 699 (2022)

Research on the Improvement of Spectra Classification Performance With the High-Performance Hybrid Deep Learning Network

Zhong-bao LIU^1、* and Jie WANG^{2、2; *;}

Author Affiliations

¹1. School of Information Science, Beijing Language and Culture University, Beijing 100083, China

²2. Xinjiang Astronomical Observatory, Chinese Academy of Sciences, Urumqi 830011, China

show less

DOI: 10.3964/j.issn.1000-0593(2022)03-0699-05 Cite this Article

Zhong-bao LIU, Jie WANG. Research on the Improvement of Spectra Classification Performance With the High-Performance Hybrid Deep Learning Network[J]. Spectroscopy and Spectral Analysis, 2022, 42(3): 699 Copy Citation Text

show less

Fig. 1. The structure of BERT-CNN

Download full size | View in the Article

Stellar Subclass Type	K1	K3	K5	K7
SNRs	(60, 65)	(60, 65)	(60, 65)	(60, 65)
Number	1115	959	850	317

Table 1. The dataset of K stars

View in the Article

Stellar Subclass Type	G0	G2	G5
SNRs	(55, 65)	(60, 65)	(40, 70)
Number	949	992	600

Table 1. The dataset of G stars

View in the Article

Stellar Subclass Type	F2	F5	F9
SNRs	(50, 65)	(65, 70)	(75, 80)
Number	1 915	1 671	1 535

Table 1. The dataset of F stars

View in the Article

参数	CNN	BERT-CNN
batch_size	128	32
learning_rate	1×10^-3	1×10^-3
hidden_units	128	256
dropout	0.5	0.5

Table 2. The parameters of CNN and BERT-CNN

View in the Article

Training Size	Test Size	P	R	F1
30%(762)	70%(1 779)	0.842 2	0.873 1	0.857 3
40%(1 016)	60%(1 525)	0.892 5	0.905 5	0.899 0
50%(1 271)	50%(1 270)	0.916 1	0.933 3	0.924 6
60%(1 525)	40%(1 016)	0.937 8	0.945 6	0.941 7
70%(1 779)	30%(762)	0.962 0	0.965 9	0.963 9

Table 3. The experimental results of BERT-CNN on the G-type dataset

View in the Article

Training Size	Test Size	P	R	F1
30%(1 536)	70%(3 585)	0.847 1	0.866 5	0.856 7
40%(2 048)	60%(3 073)	0.875 6	0.890 0	0.882 7
50%(2 561)	50%(2 560)	0.910 1	0.927 9	0.918 9
60%(3 073)	40%(2 048)	0.947 0	0.926 6	0.936 7
70%(3 585)	30%(1 536)	0.956 5	0.970 8	0.965 6

Table 3. The experimental results of BERT-CNN on the F-type dataset

View in the Article

Training Size	Test Size	P	R	F1
30%(972)	70%(2 269)	0.881 4	0.870 5	0.875 9
40%(1 296)	60%(1 945)	0.906 4	0.897 8	0.902 1
50%(1 621)	50%(1 620)	0.935 1	0.927 1	0.931 1
60%(1 945)	40%(1 296)	0.943 6	0.952 7	0.948 1
70%(2 269)	30%(972)	0.969 6	0.980 9	0.975 2

Table 3. The experimental results of BERT-CNN on the K-type dataset

View in the Article

Stellar Type	Training Size	Test Size	SVM	CNN	BERT-CNN
K	70%(2 269)	30%(972)	0.849 8	0.880 7	0.931 1
F	70%(3 585)	30%(1 536)	0.831 4	0.889 3	0.910 8
G	70%(1 779)	30%(762)	0.871 4	0.904 2	0.923 9
Average classification accuracy			0.850 9	0.891 4	0.921 9

Table 4. Comparison of experimental results

Download Citation

Save the article for my favorites

Paper Information