(3dcnn)Deep Feature Extraction and Classification of Hyperspectral Images Based on Convolutional Neural Networks

16_tgrs

Posted by Sun on March 18, 2020

16tgrs(3dcnn)Deep Feature Extraction and Classification of Hyperspectral Images Based on Convolutional Neural Networks

整体流程

3D-CNN

在第$i$层第$j$个特征图的位置$(x,y,z)$处神经元的值$v_{i,j}^{xyz}$

image-20200312104532915

image-20200312103201477

where m indexes the feature map in the $(i − 1)$th layer con- nected to the current (jth) feature map, and Pi and Qi are the height and the width of the spatial convolution kernel. Ri is the size of the kernel along toward spectral dimension, $w^{pqr}_{ijm}$ is the value of position$ (p, q, r)$ connected to the mth feature map, and bijis the bias of the jth feature map in the ith layer.

L2 Regularization of CNN + Dropout + ReLu– 防止过拟合

image-20200312102837430

加入到Loss中鼓励参数尽量小

Dropout方法防止复杂的共同适应(co-adaptation).

Dropout和ReLU共同作用使很多神经元的输出变为0, 实现强大的基于稀疏的正则化深度网络, 解决HSI中的过拟合问题

VIRTUAL SAMPLE ENHANCED CNN–扩充样本方法

CNN需要很多的样本来训练大量的参数, 但HSI样本标记获取困难, 即样本数量少. 所以需要虚拟样本方法来扩充样本

2种扩充样本的方法

Although there are many other methods that can generate the virtual samples, the changing radiation and mixture-based methods are simple yet effective ways.

  1. Changing Radiation-Based Virtual Samples image-20200312112056776

    New virtual sample $y_n$ is obtained by multiplying a random factor and adding random noise to a training sample $x_m$

    The training sample xmis a cube extracted from the hyper- spectral cube

    αm indicates the disturbance of light intensity

    β controls the weight of the random Gaussian noise n

  2. Mixture-Based Virtual Samples Because of the long distance between the object and the sensor, mixture is very common in remote sensing.??? Inspired by the phenomenon, it is possible to generate a virtual sample yk from two given samples of the same class with proper ratios

  3. image-20200312153238697

    $x_i$and $x_j$are two training samples from the same class

Feature analyse 特征表示/特征分析

1D-CNN convolution kernal in HSI

image-20200312155930922

image-20200312155215988

第一层6个1*8卷积核, (a)随机初始化的权重, (b)学习到的权重

image-20200312155440092

(a)学习到的第二层卷积核权重, 12个卷积核, 每个卷积核包含6*7个权重 (b)学习到的第三层卷积核权重, 24个卷积核, 每个卷积核包含12*8个权重

Q: 为什么6个1*8的核

A: 1个卷积核产生一种feature maps, 这个卷积核会在feature map特定不同位置内容作出反应,所以这个网络的结构应该在每层有如下特征图数量: 1 –> 6 –> 12 –> 24, 因此权值共享的意思即每个核在同一个特征图上用同样的权值,而每个核又有不一样的权重 因此可以说, CNN的学习过程: 更新卷积核参数(weights),就相当于是一直在更新所提取到的图像特征,以得到可以把图像正确分类的最合适的特征们。(一句话:更新weights以得到可以把图像正确分类的特征。)

经过convolution和pooling提取特征后的波段信息:

image-20200312154506275

2D-CNN convolution kernal in HSI

image-20200312182914800

image-20200312182226368

第一个卷积层对应的卷积核, 32个4*4的卷积核, 像素代表权重值,

(a)Pavia数据集随机出初始化的权重 (b)学习到的权重系数

特征提取

image-20200312190736709

Result

image-20200312231340782