Hyperspectral image classification based on improved multi-scale residual network structure

Hyperspectral image (HSI) is a kind of special remote sensing image, which provides rich spatial information as well as spectral information of ground objects. 3D-CNN can extract the spectral and spatial features of hyperspectral image based on this characteristic of hyperspectral image. Firstly, th...

Full description

Saved in:
Bibliographic Details
Published in2021 2nd International Symposium on Computer Engineering and Intelligent Communications (ISCEIC) pp. 377 - 382
Main Authors Guan, Li, Han, Yubing, Zhang, Pandong
Format Conference Proceeding
LanguageEnglish
Published IEEE 01.08.2021
Subjects
3D-CNN
Data mining
Deep learning
Feature extraction
Fitting
Hyperspectral image
Hyperspectral image classification
Joint features of space and spectrum
Memory management
Multi-scale residual block
Support vector machines
Training
Online AccessGet full text

Cover

More Information
Summary:Hyperspectral image (HSI) is a kind of special remote sensing image, which provides rich spatial information as well as spectral information of ground objects. 3D-CNN can extract the spectral and spatial features of hyperspectral image based on this characteristic of hyperspectral image. Firstly, the hyperspectral image data were normalized to accelerate the convergence of the network in the training. Then, a three-dimensional multi-scale residual block similar to Resnet block is designed in the network, and BN (batch normalization) layer is added to alleviate over fitting. Finally, a softmax layer outputs the classification results. The experimental results were compared with SVM and several mainstream CNN methods. In the Indian Pines dataset, compared with the performance of second model, the overall classification accuracy is increased by 1.29%, and the model parameters are around one third of the of second model; in the Pavia University dataset, the overall classification accuracy is increased by 2.1%, and the model parameters are also about one third of the performance the second model. The effects of skip-connection, pixel block size, and different spectral step of first convolution layer are also discussed. Experiments show that the network model proposed in this paper can extract better classification features and has less parameters than the traditional hyperspectral image classification algorithm, and make the hyperspectral remote sensing image classification more accurate and efficient.
DOI:10.1109/ISCEIC53685.2021.00085