You are on page 1of 16

Recurrent 3D CNN for Hyperspectral Remote

Sensing image classification

Xiaofei Yang
xiaofei.hitsz@gmail.com
School of Computer Science and Technology
Shenzhen Graduate School
Harbin Institute of Technology
Hyperspectral Remote Sensing image

3D Convolutional Neural Networks

3D Recurrent Convolutional Neural Networks

2 Harbin Institute of Technology Shenzhen Graduate School 23/8/3


Hyperspectral Remote Sensing image
• Hyperspectral imagery is typically collected (and represented) as a data cube
with spatial information collected in the X-Y plane, and spectral information
represented in the Z-direction.

3 Harbin Institute of Technology Shenzhen Graduate School 23/8/3


Hyperspectral Remote Sensing image

3D Convolutional Neural Networks

3D Recurrent Convolutional Neural Networks

4 Harbin Institute of Technology Shenzhen Graduate School 23/8/3


3D CNN

2D convolution

Shuiwang Ji, Wei Xu, Ming Yang and Kai Yu, D convolutional neural
networks for human action recognition, Pattern Analysis and Machine 3D convolution
Intelligence, IEEE Transactions on  (Volume:35 ,  Issue: 1 ), 2013
5 Harbin Institute of Technology Shenzhen Graduate School 23/8/3
3D CNN

3*3*5 2*2 3*3*5 2*2 50*4

Input
145*145*200

6 Harbin Institute of Technology Shenzhen Graduate School 23/8/3


Challenge
1. Hyperspectral Remote Sensing image

7 Harbin Institute of Technology Shenzhen Graduate School 23/8/3


Challenge
2. The parameters of ( W , b )

8 Harbin Institute of Technology Shenzhen Graduate School 23/8/3


Hyperspectral Remote Sensing image

3D Convolutional Neural Networks

3D Recurrent Convolutional Neural Networks

9 Harbin Institute of Technology Shenzhen Graduate School 23/8/3


基于 3D R-CNN 的高光谱影像分类方法

10 Harbin Institute of Technology Shenzhen Graduate School 23/8/3


基于 3D R-CNN 的高光谱影像分类方法

11 Harbin Institute of Technology Shenzhen Graduate School 23/8/3


3D R-CNN

12 Harbin Institute of Technology Shenzhen Graduate School 23/8/3


3D R-CNN

f f

fƕf
3D R-CNN

14 Harbin Institute of Technology Shenzhen Graduate School 23/8/3


Dataset :

•Indian Pines : consists of 145\times 145 pixels and 224 bands


•Salinas scene : consists of 512\times 217 pixels and 204 bands
•Salinas-A scene:consists of 83\times 86 pixels and 224 bands
•Pavia Centre scene: consists of 1096\times 715 pixels and 102 bands
•Pavia University scene:consists of 610\times 340 pixels and 103 bands
•Kennedy Space Center : consists of 512\times 614 pixels and 111 bands
•Botswana: consists of 1476\times 256 pixels and 145 bands

15 Harbin Institute of Technology Shenzhen Graduate School 23/8/3


Result : Indian Pines ( Accuracy )

SVM 0.90
GCK 0.94
3D-CNN 0.34
Ours ——

16 Harbin Institute of Technology Shenzhen Graduate School 23/8/3

You might also like