Hyperspectral Image Classification Based on Multi-Scale Feature Fusion Residual Network

Ziqing Deng; Yang Wang; Bing Zhang; Zhao Ding; Lifeng Bian; Chen Yang

doi:10.3788/LOP202259.1810014

Laser & Optoelectronics Progress, Volume. 59, Issue 18, 1810014(2022)

Hyperspectral Image Classification Based on Multi-Scale Feature Fusion Residual Network

Ziqing Deng¹, Yang Wang¹, Bing Zhang¹, Zhao Ding¹, Lifeng Bian², and Chen Yang^1、*

¹Engineering Research Center of Semiconductor Power Device Reliability, Ministry of Education, College of Big Data and Information Engineering, Guizhou University, Guiyang 550025, Guizhou , China

²Suzhou Institute of Nano-Tech and Nano-Bionics, Chinese Academy of Sciences, Suzhou 215123, Jiangsu , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(13)

Fig. 1. Structure of MFFI network

Download full size

Fig. 2. MFFI block

Download full size

Fig. 3. 3D convolution

Download full size

Fig. 4. Basic residual structure

Download full size

Fig. 5. Classification maps for IN dataset. (a) Picture of samples; (b) ground-truth label; (c) classification map of SVM; (d) classification map of 3D CNN; (e) classification map of SSRN; (f) classification map of proposed network

Download full size

Fig. 6. Classification maps for SA dataset. (a) Picture of samples; (b) ground-truth label; (c) classification map of SVM; (d) classification map of 3D CNN; (e) classification map of SSRN; (f) classification map of proposed network

Download full size

Fig. 7. Classification maps for UP dataset. (a) Picture of samples; (b) ground-truth label; (c) classification map of SVM; (d) classification map of 3D CNN; (e) classification map of SSRN; (f) classification map of proposed network

Download full size

Table 1. Classification results of three datasets
View table
Table 1. Classification results of three datasets
Dataset OA /% AA /% K
IN 99.60 99.55 0.9955
UP 99.85 99.80 0.9980
SA 99.95 99.89 0.9995

Table 2. Average accuracy of three datasets with different number of blocks
View table
Table 2. Average accuracy of three datasets with different number of blocks
MFFI number OA /% AA /% K
One MFFI 98.88 99.11 0.9869
Two MFFIs 98.42 96.85 0.9821
Three MFFIs 99.61 99.71 0.9955
Four MFFIs 99.64 99.64 0.9958
Five MFFIs 99.59 99.58 0.9959

Table 3. Average accuracy and average training time of three datasets with different input sizes
View table
Table 3. Average accuracy and average training time of three datasets with different input sizes
Input size OA /% AA /% K Training time /（s·epoch^-1）
7×7×N 99.64 99.64 0.9958 43
9×9×N 99.75 99.74 0.9971 56
11×11×N 99.77 99.75 0.9973 71
13×13×N 99.78 99.69 0.9974 132

Table 4. Classification result and training time of different networks on IN dataset
View table
Table 4. Classification result and training time of different networks on IN dataset
Results SVM SAE 3D CNN SSRN Proposed
OA /% 81.67 92.99 97.08 99.19 99.61
AA /% 79.84 89.76 95.09 98.93 99.55
K 0.7876 0.9218 0.9674 0.9907 0.9955
Training time /（s·epoch^-1） 6 20 34 92.8 41

Table 5. Classification result and training time of different networks on UP dataset
View table
Table 5. Classification result and training time of different networks on UP dataset
Results SVM SAE 3D CNN SSRN Proposed
OA /% 90.08 86.10 98.85 99.79 99.82
AA /% 92.99 76.09 98.40 99.66 99.79
K 0.8721 0.8109 0.9847 0.9972 0.9976
Training time /（s·epoch^-1） 4 25 38 61.2 50

Table 6. Classification result and training time of different networks on SA dataset
View table
Table 6. Classification result and training time of different networks on SA dataset
Results SVM SAE 3D CNN SSRN Proposed
OA /% 88.16 92.01 94.19 99.46 99.95
AA /% 93.29 96.56 95.65 99.77 99.89
K 0.8680 0.9110 0.9353 0.9940 0.9995
Training time /（s·epoch^-1） 6 46 72 302 132

Tools

Get Citation

Copy Citation Text

Ziqing Deng, Yang Wang, Bing Zhang, Zhao Ding, Lifeng Bian, Chen Yang. Hyperspectral Image Classification Based on Multi-Scale Feature Fusion Residual Network[J]. Laser & Optoelectronics Progress, 2022, 59(18): 1810014

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Image Processing

Received: Jun. 15, 2021

Accepted: Aug. 10, 2021

Published Online: Aug. 29, 2022

The Author Email: Yang Chen (eliot.c.yang@163.com)

DOI:10.3788/LOP202259.1810014

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Classification results of three datasets

Table 1. Classification results of three datasets

Table 2. Average accuracy of three datasets with different number of blocks

Table 2. Average accuracy of three datasets with different number of blocks

Table 3. Average accuracy and average training time of three datasets with different input sizes

Table 3. Average accuracy and average training time of three datasets with different input sizes

Table 4. Classification result and training time of different networks on IN dataset

Table 4. Classification result and training time of different networks on IN dataset

Table 5. Classification result and training time of different networks on UP dataset

Table 5. Classification result and training time of different networks on UP dataset

Table 6. Classification result and training time of different networks on SA dataset

Table 6. Classification result and training time of different networks on SA dataset