Neural architecture search combined with efficient attention for hyperspectral image classification

Haisong CHEN; Kang ZHANG; Haoran LÜ; Aili WANG; Haibin WU

doi:10.37188/CJLCD.2024-0254

Chinese Journal of Liquid Crystals and Displays, Volume. 40, Issue 4, 630(2025)

Neural architecture search combined with efficient attention for hyperspectral image classification

Haisong CHEN¹, Kang ZHANG², Haoran LÜ², Aili WANG^2、*, and Haibin WU²

Author Affiliations

¹School of Integrated Circuit，Shenzhen Polytechnic University，Shenzhen 518055，China

²Heilongjiang Province Key Laboratory of Laser Spectroscopy Technology and Application，Harbin University of Science and Technology，Harbin 150080，China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(20)

Fig. 1. Network structure of EA-NAS

Download full size

View in Article

Fig. 2. Operators of attention

Download full size

View in Article

Fig. 3. Modular convolution operators

Download full size

View in Article

Fig. 4. Search process

Download full size

View in Article

Fig. 5. Classification results of KSC dataset

Download full size

View in Article

Fig. 6. Classification results of PU dataset

Download full size

View in Article

Fig. 7. Optimal structure of cell on PU dataset

Download full size

View in Article

Fig. 8. Optimal structure of cell on KSC dataset

Fig. 9. Loss curves

Fig. 10. Accurcy curves

Download full size

View in Article

Fig. 11. Feature map of no-moducular convolution operator t-SNE

Download full size

View in Article

Fig. 12. Feature map of modular convolution operator t-SNE

Download full size

View in Article

Table 1. Details of KSC dataset
View table
View in Article
Table 1. Details of KSC dataset
No. Class Color Sample
Total 5 211
1 Scrub 761
2 Willow 243
3 Palm 256
4 Pine 252
5 Broadleaf 161
6 Hardwood 229
7 Swap 105
8 Graminoid 431
9 Spartina 520
10 Cattail 404
11 Salt 419
12 Mud 503
13 Water 927

Table 2. Details of PU dataset
View table
View in Article
Table 2. Details of PU dataset
No. Class Color Sample numbers
Total 42 776
1 Asphalt 6 631
2 Meadows 18 649
3 Gravel 2 099
4 Trees 3 064
5 Painted metal sheets 1 345
6 Bare Soil 5 029
7 Bitumen 1 330
8 Self-Blocking Bricks 3 682
9 Shadows 947

Table 3. Parameter settings of each operator in the search space

View table

View in Article

Table 3. Parameter settings of each operator in the search space

序号	算子名称	参数设置
1	CBAM注意力算子	无
2	FusedMB_conv_3_3	Conv（3×3）-SE-Conv（1×1）
3	FusedMB_conv_3_5	Conv（3×5）-SE-Conv（1×1）
4	FusedMB_conv_3_7	Conv（3×7）-SE-Conv（1×1）
5	MB_conv_3_3	Conv（1×1）-DSepConv（3×3）-SE-Conv（1×1）
6	MB_conv_3_5	Conv（1×1）-DSepConv（3×3）-SE-Conv（1×1）
7	MB_conv_3_7	Conv（1×1）-DSepConv（3×3）-SE-Conv（1×1）

Table 4. Comparison of classification accuracy in KSC dataset

View table

View in Article

Table 4. Comparison of classification accuracy in KSC dataset

No.	RBF-SVM	3D-CNN	ResNet	3D Auto-CNN	HNAS	EA-NAS
1	92.71±0.68	94.07±1.31	97.04±2.66	92.50±7.14	99.43±1.22	99.96±0.06
2	84.62±4.02	92.12±3.75	86.83±7.86	96.31±4.82	93.58±1.63	98.46±3.06
3	73.64±7.94	89.73±7.54	85.53±8.07	94.07±6.10	95.29±6.74	100±0.00
4	54.33±8.62	80.75±8.36	78.67±9.02	95.92±5.32	95.76±2.19	96.95±4.43
5	61.46±10.33	80.86±4.17	67.70±2.14	97.04±2.34	96.47±4.24	96.26±3.57
6	65.51±8.21	84.74±7.33	93.52±6.31	98.88±2.73	93.10±7.48	99.82±0.13
7	76.52±3.37	93.56±10.21	95.12±7.70	94.54±10.91	98.23±0.30	99.54±0.91
8	86.20±5.71	94.78±1.07	95.09±3.48	94.53±6.91	98.37±1.62	99.60±0.52
9	88.71±2.03	96.76±1.21	97.14±2.21	84.77±10.91	99.86±0.34	100±0.00
10	96.72±4.04	98.91±0.23	99.51±0.94	98.85±1.61	99.66±1.00	99.40±1.19
11	96.31±1.62	99.32±0.76	99.81±0.54	100±0.00	99.71±0.77	99.53±0.67
12	93.56±2.49	98.12±1.86	96.33±2.12	97.96±5.25	99.18±1.26	99.87±0.88
13	99.88±0.02	99.64±0.37	98.83±1.98	99.18±1.99	100±0.00	100±0.00
OA/%	87.88±0.97	94.89±0.25	93.72±2.00	94.61±1.73	97.90±0.87	99.50±0.21
AA/%	82.92±2.09	92.37±0.48	91.63±2.68	95.26±0.92	96.49±0.26	99.20±0.42
K×100	86.27±1.35	94.66±0.61	93.00±2.23	93.98±1.94	96.89±0.97	99.44±0.23

Table 5. Comparison of classification accuracy in PU dataset

View table

View in Article

Table 5. Comparison of classification accuracy in PU dataset

No.	RBF-SVM	3D-CNN	ResNet	3D Auto-CNN	HNAS	EA-NAS
1	81.14±4.98	94.94±1.41	96.94±1.81	92.87±4.48	92.10±4.42	96.92±1.12
2	83.93±1.31	96.97±0.75	98.92±1.41	98.82±0.79	98.79±0.65	98.07±1.67
3	57.56±15.11	73.81±11.02	92.76±4.85	91.17±5.98	93.76±2.34	95.85±4.94
4	94.37±3.19	75.44±12.54	99.75±0.15	91.28±10.16	92.73±4.34	99.52±0.42
5	95.44±3.10	98.73±0.14	99.80±0.21	93.96±5.94	92.16±5.87	99.56±0.48
6	81.02±5.91	80.32±5.43	98.68±1.05	99.09±0.94	99.10±0.81	98.05±0.91
7	69.77±10.94	68.62±14.15	96.04±4.34	88.71±5.92	90.94±4.75	96.40±1.89
8	71.73±6.01	80.91±3.22	89.52±5.19	89.07±4.78	89.83±4.24	91.24±2.67
9	99.86±0.05	97.23±5.03	98.98±0.84	85.03±7.22	86.92±6.43	93.55±1.17
OA/%	82.16±2.01	86.85±3.07	97.36±0.85	94.85±0.89	95.11±0.87	96.81±0.26
AA/%	78.09±4.57	85.77±4.02	96.82±0.75	92.18±1.64	92.89±0.77	96.35±0.32
K×100	75.21±4.09	84.16±4.06	96.49±1.13	91.84±4.74	93.81±0.77	96.03±0.34

Table 6. Comparison of parameter count and runtime of different methods on two datasets
View table
View in Article
Table 6. Comparison of parameter count and runtime of different methods on two datasets
Methods KSC PU
参数量/M Train time/min Test time/s 参数量/M Train time/min Test time/s
3D Auto-CNN 0.12 19.41 7.72 0.19 11.22 5.79
HNAS 2.70 31.63 8.94 2.73 15.35 8.89
EA-NAS 1.33 11.57 9.47 1.52 12.77 12.48

Table 7. Ablation experiment
View table
View in Article
Table 7. Ablation experiment
Attention Convolution Block Poly OA/% AA/% K×100
√ √ 92.30 89.78 89.96
√ √ 96.34 96.23 96.33
√ √ 96.33 96.54 96.84
√ √ √ 96.81 96.35 96.03

Table 8. Modular convolutional operator ablation experiment
View table
View in Article
Table 8. Modular convolutional operator ablation experiment
SE卷积算子 CBAM卷积算子 FusedMB卷积算子 MB卷积算子 OA/% AA/% K×100
√ √ 94.61 95.53 95.47
√ √ 92.30 89.78 89.96
√ √ 96.33 96.54 96.84
√ √ √ √ 96.81 96.35 96.03

Tools

Get Citation

Copy Citation Text

Haisong CHEN, Kang ZHANG, Haoran LÜ, Aili WANG, Haibin WU. Neural architecture search combined with efficient attention for hyperspectral image classification[J]. Chinese Journal of Liquid Crystals and Displays, 2025, 40(4): 630

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category:

Received: Aug. 28, 2024

Accepted: --

Published Online: May. 21, 2025

The Author Email: Aili WANG (aili925@hrbust.edu.cn)

DOI:10.37188/CJLCD.2024-0254

Topics

Table 1. Details of KSC dataset

Table 1. Details of KSC dataset

Table 2. Details of PU dataset

Table 2. Details of PU dataset

Table 3. Parameter settings of each operator in the search space

Table 3. Parameter settings of each operator in the search space

Table 4. Comparison of classification accuracy in KSC dataset

Table 4. Comparison of classification accuracy in KSC dataset

Table 5. Comparison of classification accuracy in PU dataset

Table 5. Comparison of classification accuracy in PU dataset

Table 6. Comparison of parameter count and runtime of different methods on two datasets

Table 6. Comparison of parameter count and runtime of different methods on two datasets

Table 7. Ablation experiment

Table 7. Ablation experiment

Table 8. Modular convolutional operator ablation experiment

Table 8. Modular convolutional operator ablation experiment