Remote sensing image scene classification based on multilayer feature context encoding network

Ruo-Yao LI; Bo ZHANG; Bin WANG

doi:10.11972/j.issn.1001-9014.2021.04.012

Journal of Infrared and Millimeter Waves, Volume. 40, Issue 4, 530(2021)

Remote sensing image scene classification based on multilayer feature context encoding network

Ruo-Yao LI^1,2, Bo ZHANG^1,2, and Bin WANG^1,2、*

Author Affiliations

¹Key Laboratory for Information Science of Electromagnetic Waves （MoE）， Fudan University， Shanghai 200433， China

²Research Center of Smart Networks and Systems， School of Information Science and Technology， Fudan University， Shanghai 200433， China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(11)

Fig. 1. The illustration of the architecture of DenseNet

Download full size

View in Article

Fig. 2. The framework of the proposed MFCE network Note： $⊙$ and $↑$ denote the channel concatenation operation and the spatial up-sampling operation， respectively

Download full size

View in Article

Fig. 3. Samples of remote sensing images (a) AID dataset，(b) NWPU-RESISC45 dataset

Download full size

View in Article

Fig. 4. Test accuracy with MFCE network and Fine-tuned DenseNet-121 (a) AID dataset, (b) NWPU-RESISC45 dataset

Download full size

View in Article

Fig. 5. Visual comparison of heatmaps among MFCE andFine-tuned DenseNet-121 for NWPU-RESISC45 dataset Note: (d-f) heatmaps of the baseline，and (g-i) MFCE network

Download full size

View in Article

Table 1. OA of different methods on AID dataset with different training ratios

View table

View in Article

Table 1. OA of different methods on AID dataset with different training ratios

Method	OA
Method	Tr=20%	Tr=50%
VGG-VD-16^［17］	86.59 $\pm$ 0.29	89.64 $\pm$ 0.36
Fine-tuned DenseNet-121	94.75 $\pm$ 0.18	96.56 $\pm$ 0.17
FACNN^［12］	-	95.45 $\pm$ 0.11
D-CNN with VGGNet-16^［7］	90.82 $\pm$ 0.16	96.89 $\pm$ 0.10
VGG-VD16+MSCP+MRA^［11］	92.21 $\pm$ 0.17	96.56 $\pm$ 0.18
MFCE （ours）	95.51 $\pm$ 0.09	97.14 $\pm$ 0.19

Table 2. OA of different methods on NWPU-RESISC45 dataset with different training ratios

View table

View in Article

Table 2. OA of different methods on NWPU-RESISC45 dataset with different training ratios

Method	OA
Method	Tr=10%	Tr=20%
VGGNet-16^［18］	87.15 $\pm$ 0.45	90.36 $\pm$ 0.18
Fine-tuned DenseNet-121	91.56 $\pm$ 0.21	93.72 $\pm$ 0.20
FACNN^［12］	-	-
VGG-VD16+MSCP+MRA^［11］	88.07 $\pm$ 0.18	90.81 $\pm$ 0.13
D-CNN with VGGNet-16^［7］	89.22 $\pm$ 0.50	91.89 $\pm$ 0.22
MFCE （ours）	92.42 $\pm$ 0.20	94.40 $\pm$ 0.09

Table 3. Results of MFCE network adopting different levels of multiscale pooling on AID dataset and NWPU-RESISC45 dataset
View table
View in Article
Table 3. Results of MFCE network adopting different levels of multiscale pooling on AID dataset and NWPU-RESISC45 dataset
Methods OA
AID （Tr=20%） NWPU-RESISC45 （Tr=10%）
MFCE （2， 4， 6） 95.16 $\pm$ 0.20 92.17 $\pm$ 0.28
MFCE （2， 4， 6， 8） 95.51 $\pm$ 0.09 92.42 $\pm$ 0.20

Table 4. Comparison of different methods on AID dataset and NWPU-RESISC45 dataset
View table
View in Article
Table 4. Comparison of different methods on AID dataset and NWPU-RESISC45 dataset
Method OA
AID （Tr=20%） NWPU-RESISC45 （Tr=10%）
Fine-tuned DenseNet-121 94.75 $\pm$ 0.18 91.56 $\pm$ 0.21
MFCE without Context Encoding 94.92 $\pm$ 0.19 91.52 $\pm$ 0.30
MFCE 95.51 $\pm$ 0.09 92.42 $\pm$ 0.20

Table 5. Results of MCE module combined with different backbones and baselines on AID dataset
View table
View in Article
Table 5. Results of MCE module combined with different backbones and baselines on AID dataset
Method OA
Fine-tuned VGGNet-16 90.19 $\pm$ 0.38
Fine-tuned ResNet-18 93.10 $\pm$ 0.35
Fine-tuned DenseNet-121 94.75 $\pm$ 0.18
VGGNet-16+MCE （ours） 91.57 $\pm$ 0.26
ResNet-18+MCE （ours） 94.08 $\pm$ 0.20
MFCE （ours） 95.51 $\pm$ 0.09

Table 6. Parameters and MACs of different networks

View table

View in Article

Table 6. Parameters and MACs of different networks

Methods	Parameters	MACs
VGGNet-16^［5］	138.36M	15.48G
ResNet-18^［19］	11.69M	1.82G
DenseNet-121^［13］	7.98M	2.87G
VGGNet-16+MCE （ours）	15.52M	20.10G
ResNet-18+MCE （ours）	11.98M	2.42G
MFCE （ours）	10.14M	3.91G

Tools

Get Citation

Copy Citation Text

Ruo-Yao LI, Bo ZHANG, Bin WANG. Remote sensing image scene classification based on multilayer feature context encoding network[J]. Journal of Infrared and Millimeter Waves, 2021, 40(4): 530

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Research Articles

Received: Oct. 9, 2020

Accepted: --

Published Online: Sep. 9, 2021

The Author Email: Bin WANG (wangbin@fudan.edu.cn)

DOI:10.11972/j.issn.1001-9014.2021.04.012

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. OA of different methods on AID dataset with different training ratios

Table 1. OA of different methods on AID dataset with different training ratios

Table 2. OA of different methods on NWPU-RESISC45 dataset with different training ratios

Table 2. OA of different methods on NWPU-RESISC45 dataset with different training ratios

Table 3. Results of MFCE network adopting different levels of multiscale pooling on AID dataset and NWPU-RESISC45 dataset

Table 3. Results of MFCE network adopting different levels of multiscale pooling on AID dataset and NWPU-RESISC45 dataset

Table 4. Comparison of different methods on AID dataset and NWPU-RESISC45 dataset

Table 4. Comparison of different methods on AID dataset and NWPU-RESISC45 dataset

Table 5. Results of MCE module combined with different backbones and baselines on AID dataset

Table 5. Results of MCE module combined with different backbones and baselines on AID dataset

Table 6. Parameters and MACs of different networks

Table 6. Parameters and MACs of different networks