生成式对抗网络的高光谱遥感图像分类方法研究
Research on classification method of hyperspectral remote sensing image based on Generative Adversarial Network
- 2022年26卷第2期 页码:416-430
纸质出版日期: 2022-02-07
DOI: 10.11834/jrs.20219192
引用
阅读全文PDF
扫 描 看 全 文
纸质出版日期: 2022-02-07
扫 描 看 全 文
引用
阅读全文PDF
针对基于深度学习的分类模型在训练样本较少时所遭受的潜在过拟合问题,提出一种具备过拟合抑制的生成式对抗网络分类算法,并应用于高光谱图像分类。该算法在每次迭代时,首先,依据训练样本的标签信息使判别器网络拟合训练样本的数据分布;然后对训练样本的高维特征进行均值最小化,该过程会重新更新判别器网络参数,减小参数的值和方差,以抑制过拟合;最后,将本算法应用于针对高光谱图像所设计的光谱空间分类模型进行分类。实验结果表明,在标准数据集Indian Pines和Pavia University中随机选取1%标记样本进行训练,总体分类精度分别达到了89.61%和98.79%,相比于其他现有算法有明显的提高,较表现最好的分类方法,总体分类精度分别提升了5.17%和1.38%。在Indian Pines数据集取1%标记样本,Pavia University数据集取0.1%标记样本的情况下,本文算法对过拟合的抑制效果优于几种常用的过拟合抑制算法,较表现最好的Dropout算法,总体分类精度分别提升了5.60%和3.20%。
Deep learning has strong learning ability and has become a widely studied method in the hyperspectral image classification community. However, the deep learning-based classification model requires a large number of training samples to train a good model. Overfitting will occur when the training sample is small. The accuracy of the model on the test set is lower than the accuracy on the training set. Researchers have proposed overfitting suppression methods such as weight decay and dropout to suppress overfitting. However, these methods need to work in a specific environment and have limited suppression effect on overfitting. Thus, this study proposes an overfitting suppression algorithm based on generative adversarial networks to suppress the overfitting phenomenon of the model.
First, a spatial neighborhood block for the standard dataset is constructed, and the dataset is divided into labeled, unlabeled, and test samples. Then, the labeled and unlabeled samples are sent to the generative adversarial networks for training. During input, the pixels in the neighborhood block are independently fed into the fully connected network discriminator to extract the spectral features of each pixel. Finally, the spectral features of each pixel are fused by the average pooling, and they connected to the output layer to obtain the classification result. The overfitting is caused by the large value and variance of the network parameters. Thus, the large parameter values enable the model to fit more samples. Therefore, the network is first fitted to the data by labeled samples in each iteration, and then, the optimizer is used to minimize the mean of the high-dimensional features. This process will re-update the network parameters, reduce the value and variance of the parameters, and thus suppress the overfitting.
The algorithm was applied to two standard datasets, namely, Indian Pines and Pavia University datasets. The 1% labeled samples were randomly selected for training. The overall classification accuracy rates were 89.61% and 98.79%, which were better than those of several algorithms. Compared with several commonly used overfitting suppression methods such as batch normalization, L2 regularization, and dropout, the proposed overfitting suppression algorithm obtains 5.60% and 3.20% higher results on randomly selected 1% labeled samples from the Indian Pines dataset and randomly selected 0.1% labeled samples from Pavia University dataset.
The model of generative adversarial networks designed for the characteristics of hyperspectral data can fully utilize the spectral and spatial features of hyperspectral images. The proposed overfitting suppression algorithm can significantly improve the classification performance of the model. However, the overfitting suppression effect of the algorithm is not obvious when the number of labeled samples is large. Thus, further research is needed.
高光谱图像(HSI)通常包含数百个连续波段,为区分地物类别提供了可靠的依据(
在过去的几十年中,国内外学者提出了大量的HSI分类方法。如基于支持向量机(SVM)的分类方法(
目前,抑制过拟合的方法有数据集扩增,提前停止,集成学习,批量归一化(Batch Normalization (BN),L2正则化,Dropout等。数据集扩增后可以利用更多的规则来约束模型,使模型更接近于数据的真实分布,然而在实际应用中高光谱遥感图像标签获取成本高,标记样本较少,数据集扩增方法并不可行。提前停止通过设置约束条件,使网络在测试集准确率不再提升或测试集损失不再降低时停止训练,可以防止由于训练次数过多导致模型过拟合(
本文受对抗思想的启发,提出一种基于GAN的高维特征均值最小化(HFAM)过拟合抑制算法。利用优化函数对神经网络输出层前一层特征的均值进行最小化,以减小网络参数。较小的网络参数能够有效抑制过拟合(
大量研究表明,GAN能够很好地适用于分类任务。本文针对HSI的特点设计了一种基于GAN的HSI分类模型。算法框架如
图1 本文算法框架
Fig.1 Algorithm framework
HSI通常包含数百个波段,其中某些波段可能会受噪声干扰,影响分类精度(
图2 判别器网络结构
Fig.2 Discriminator network structure
GAN包括一个判别器D和一个生成器G。D用于判别数据是否来自真实数据分布,G用于生成尽可能真实的数据去欺骗D,在理想状态下,G可以生成判别器无法判断真假的数据。该过程可以表示为
minGmaxDV(D,G)=𝔼x~pdata(x)(log D(x))+𝔼z~pz(z)(log(1-D(G(z)))) | (1) |
式中,x为真实数据,Pdata(x)为x的概率分布,D(x)为D将x判断为真的概率,z为噪声数据,P(z)为z的概率分布,D(G(z))为D将伪数据G(z)判断为真的概率,V(D,G)为损失函数。
将GAN应用于分类任务中,可以使用任何标准的深度学习分类模型,只需将G生成的伪样本添加到训练样本中,并用y=K+1标记,相应地增加判别器的输出维数K到K+1。对于未标记样本,最大化log pmodel(y<K+1| x)使其对应于K类真实数据之一。用于训练GAN的损失函数为
L=Lsupervised+Lunsupervised | (2) |
Lsupervised=𝔼x,y~pdata(x,y)log pmodel(y | x,y<K+1) | (3) |
Lunsupervised=-𝔼x~pdata(x)log(1-pmodel(y=K+1| x))-𝔼x~𝔾log pmodel(y=K+1| x) | (4) |
式中,L为总损失,Lsupervised为有监督分类的损失,Lunsupervised为无监督分类的损失。pmodcl(y | x,y<K+1)为将x分到正确类的概率,pmodel(y=K+1 | x)为将x判断为假的概率。
过拟合是导致基于深度学习的分类模型准确率难以提升的主要原因之一(
过拟合时拟合函数需要顾及每一个样本,越复杂的模型,其参数受噪声的影响越严重,当模型复杂性增加时,通常偏差减小,方差增加(
基于此,本文受对抗思想的启发,提出一种新的过拟合抑制算法HFAM,判别器网络一方面根据标记样本的标签信息拟合数据的分布;另一方面通过对标记样本的高维特征均值最小化并反向更新网络来减小网络参数的值和方差,以抑制判别器的拟合效果,抑制过拟合的过程为
θ*=argminθ1NN∑i=1(average(f(xi;θ))) | (5) |
式中,xi为训练样本,θ为网络参数,N为一次迭代的样本数量,f为模型的高维输出,本算法中代表的是平均池化层的输出,是一个一维特征向量。在本文提出的模型中,3个全连接层都使用ReLU激活函数,每一层的输出都是非负数,故f也为非负数。最小化f中特征的平均值,可以使网络参数减小,抑制拟合过程,同时可以减小网络参数的方差,使参数更加平滑。平滑的网络参数会充分考虑每个波段的特征,不太依赖某些局部特征,泛化能力更强。抑制过拟合算法流程为
输入:训练样本Xtrain={x1train,x2train,⋯,xntrain},训练样本标签Ytrain={y1train,y2train,⋯,yntrain};
测试样本Xtest={x1test,x2test,⋯,xmtest},测试样本标签Ytest={y1test,y2test,⋯,ymtest};
每次迭代的样本数量 k.
过程:
计算完成一轮训练所需的迭代次数l=n/k
repeat
for i=1,2,⋯,l do
从Xtrain中选取一次迭代的训练样本{x1train,x2train,⋯,xktrain}
for j=1,2,⋯,k do
计算xjtrain的输出f1(xjtrain;θ0)与标签yjtrain的损失Li j(yjtrain,f1(xjtrain;θ0))
end for
计算当前迭代样本的平均损失Li=1kk∑j=1(Li j)
通过优化函数最小化Li,得到参数θ1=arg minθ0(Li)
for j=1,2,⋯,k do
将xjtrain输入网络得到xjtrain的高维特征hi j=f2(xjtrain;θ1)
计算xjtrain高维特征的平均值average(hi j)
end for
计算当前迭代样本高维特征的平均值hi=1kk∑j=1average(hi j)
通过优化函数最小化hi得到新的参数θ=arg minθ1(hi)
end for
将测试样本Xtest送入网络得到Ypred,根据Ytest计算测试样本的准确率
until测试样本的准确率不再提升
输出:训练好的分类模型
本文使用的是两个具有不同光谱和空间分辨率的真实高光谱遥感数据集Indian Pines和Pavia University。
(1)Indian Pines数据集是最早用于HSI分类的测试数据,由AVIRIS传感器1992年在印第安纳州西北部农业区成像并截取145×145大小的图像进行标注得到的印度松树数据集。波长范围为0.4—2.5 μm,空间分辨率为20 m,共包含220个波段,但第104—108、150—163和第220个波段受水汽吸收的影响信噪比较低,去除这20个波段之后,保留了200个波段。该数据集共有21025个像元,其中10249个像元是有标签的,有标签像元共包括16种地物类别,最多的一类有2455个像元,最少的一类只有20个像元。Indian Pines数据集假彩色图片,真实地物类别图及各类样本数量如
图3 Indian Pines数据集
Fig.3 Indian Pines dataset
(2)Pavia University数据集是由意大利帕维亚大学的ROSIS传感器2003年在意大利帕维亚城市上空成像并截取610×340大小的图像进行标注得到的数据集。波长范围为0.43—0.86 μm,空间分辨率为1.3 m,共包含115个波段,在去掉12个受噪声影响的波段后,保留了103个波段。该数据集共有207400个像元,但其中只有42776个像元是有标签的,有标签像元共包含9种地物类别,最多的一类有18649个像元,最少的一类有947个像元。Pavia University数据集假彩色图片,真实地物类别图及各类样本数量如
图4 Pavia University数据集
Fig.4 Pavia University dataset
本实验在配备Intel i7 2600k CPU、8 G内存和Nvidia GTX 1080 GPU的PC机上运行。使用Window 7×64操作系统,Spyder开发环境,Tensorflow框架,Python编程语言。判别器网络主要包括3个全连接层和1个平均池化层,生成器网络主要包括两个反卷积层和3个卷积层。判别器和生成器都使用Adam优化器,学习率为0.0002,batch size为16。所有实验都按比例选取训练样本,并且无标记训练样本数量均为有标记训练样本的5倍,其余为测试样本。由于选取的训练样本较少,使用不同训练样本得到的实验结果差异较大,所有实验结果均为随机选取训练样本并运行10次的平均值。
为了测试本文算法的分类性能,两个数据集每类都选取1%标记样本(至少一个),与基于SVM的分类方法(
类别 | SVM | KELM | NRJSR | SADL | 3D CNNs | MCMs+2DCNN | MCNN | 本文算法 |
---|---|---|---|---|---|---|---|---|
1 | 7.50 | 64.81 | 75.88 | 40.57 | 32.61 | 71.11 | 71.11 | 63.64 |
2 | 58.03 | 66.32 | 64.13 | 63.81 | 74.37 | 76.23 | 85.49 | 92.72 |
3 | 26.34 | 50.11 | 47.23 | 58.84 | 74.58 | 77.08 | 87.20 | 88.81 |
4 | 1.77 | 33.46 | 45.58 | 56.10 | 35.44 | 94.36 | 71.74 | 70.15 |
5 | 61.42 | 67.74 | 73.01 | 80.18 | 69.36 | 65.82 | 80.81 | 95.60 |
6 | 90.46 | 90.42 | 97.05 | 90.09 | 85.62 | 88.39 | 90.25 | 95.97 |
7 | 14.58 | 58.33 | 70.87 | 99.60 | 35.71 | 100 | 92.59 | 71.43 |
8 | 96.01 | 97.00 | 95.08 | 99.13 | 91.00 | 80.93 | 39.87 | 100 |
9 | 22.78 | 90.56 | 22.35 | 96.32 | 95.00 | 90.53 | 73.68 | 100 |
10 | 19.89 | 53.75 | 44.73 | 64.67 | 66.46 | 68.21 | 76.56 | 82.73 |
11 | 87.70 | 75.34 | 77.57 | 75.43 | 84.93 | 86.55 | 92.61 | 86.65 |
12 | 18.73 | 52.01 | 43.49 | 48.30 | 57.50 | 42.25 | 75.13 | 90.12 |
13 | 89.04 | 99.43 | 98.95 | 98.90 | 95.61 | 86.44 | 84.42 | 87.23 |
14 | 96.61 | 93.15 | 97.55 | 90.93 | 92.41 | 90.37 | 96.09 | 99.58 |
15 | 16.17 | 29.60 | 56.99 | 59.87 | 69.43 | 65.29 | 81.55 | 84.44 |
16 | 75.91 | 80.32 | 94.35 | 85.96 | 93.55 | 71.30 | 64.44 | 77.27 |
OA | 63.35 | 70.48 | 71.68 | 73.70 | 78.02 | 78.62 | 84.44 | 89.61 |
AA | 48.94 | 68.90 | 69.05 | 75.54 | 72.10 | 75.60 | 78.97 | 86.65 |
k×100 | 56.64 | 66.14 | 67.37 | 70.06 | 74.85 | 74.85 | 82.19 | 88.14 |
注: 粗体为最好的结果。
类别 | SVM | KELM | NRJSR | SADL | 3D CNNs | MCMs+2DCNN | MCNN | 本文算法 |
---|---|---|---|---|---|---|---|---|
1 | 86.08 | 83.01 | 62.63 | 89.78 | 91.66 | 94.88 | 95.95 | 98.95 |
2 | 96.10 | 96.78 | 98.24 | 96.89 | 98.86 | 98.93 | 99.78 | 99.55 |
3 | 59.03 | 62.63 | 76.85 | 71.36 | 82.99 | 92.10 | 96.01 | 94.16 |
4 | 83.32 | 84.36 | 77.26 | 93.17 | 89.72 | 77.16 | 91.67 | 99.83 |
5 | 98.19 | 98.81 | 99.96 | 99.73 | 98.59 | 98.33 | 99.85 | 100 |
6 | 58.15 | 55.91 | 79.23 | 89.39 | 88.77 | 99.73 | 97.91 | 99.00 |
7 | 55.53 | 69.61 | 92.57 | 73.02 | 64.74 | 99.35 | 94.47 | 100 |
8 | 83.24 | 82.64 | 88.75 | 73.86 | 91.06 | 98.15 | 97.51 | 94.98 |
9 | 95.45 | 92.15 | 54.57 | 99.89 | 84.48 | 42.13 | 80.17 | 100 |
OA | 85.03 | 85.18 | 86.03 | 90.82 | 93.07 | 95.17 | 97.41 | 98.79 |
AA | 79.45 | 80.66 | 81.12 | 87.45 | 87.87 | 88.97 | 94.91 | 98.50 |
k×100 | 79.77 | 79.95 | 81.31 | 87.83 | 90.75 | 93.60 | 96.56 | 98.40 |
注: 粗体为最好的结果。
为了验证本文算法的鲁棒性,用两个具有不同波段数目,不同光谱分辨率和空间分辨率的标准数据集进行实验。Indian Pines数据集分别随机选取0.5%、1%、2%、4%训练样本。由于Pavia University数据集样本数量多,且光谱质量较高,Pavia University数据集分别随机选取0.1%、0.4%、0.7%、1%训练样本。Indian Pines数据集总体分类精度分别为81.00%、89.61%、95.42%、97.64%,Pavia University数据集总体分类精度分别为88.53%、96.75%、97.54%、98.79%,均达到了理想水平。Indian Pines数据集取不同比例训练样本的分类准确率如
图5 Indian Pines数据集分类准确率
Fig.5 Classification accuracies for Indian Pines dataset
图6 Pavia University数据集分类准确率
Fig.6 Classification accuracies for Pavia University dataset
图7 Indian Pines数据集不同比例标记样本分类结果图
Fig.7 Classification maps provided by different proportions of labeled samples on the Indian Pines dataset
图8 Pavia University数据集不同比例标记样本分类结果图
Fig.8 Classification maps provided by different proportions of labeled samples on the Pavia University dataset
HSI的空间特征表示某个像元与其周围像元很有可能是同一类地物,因此,合理利用空间特征将有助于提升分类准确率。为了验证本文所提出的模型对空间特征提取的有效性,两个数据集都取1%标记样本,采用不同大小的邻域块分别进行实验。Indian Pines数据集分类准确率如
图9 Indian Pines数据集分类准确率
Fig.9 Classification accuracies for Indian Pines dataset
图10 Pavia University数据集分类准确率
Fig.10 Classification accuracies for Pavia University dataset
在少量训练样本的情况下,基于深度学习的分类模型通常会遭受潜在的过拟合影响,导致模型在训练集上的分类准确率很高,而在测试集上的分类准确率不够理想。合理减小参数的值和方差可以有效缓解过拟合(
图11 不同方法参数的变化情况
Fig.11 Variation of parameters in different methods
为了验证所提出过拟合抑制算法HFAM的有效性,Indian Pines数据集随机选取1%标记样本,Pavia University数据集随机选取0.1%标记样本,分别使用BN、L2正则化、Dropout以及HFAM进行过拟合抑制实验。Pavia University数据集分类准确率如
类别 | 数量 | GAN | GAN+BN | GAN+L2 | GAN+Dropout | GAN + HFAM |
---|---|---|---|---|---|---|
1 | 6631 | 86.15 | 87.32 | 88.79 | 89.82 | 88.62 |
2 | 18649 | 94.73 | 94.72 | 92.97 | 90.65 | 95.60 |
3 | 2099 | 73.63 | 73.09 | 71.26 | 74.30 | 87.94 |
4 | 3064 | 71.75 | 66.39 | 80.31 | 76.89 | 87.75 |
5 | 1345 | 100 | 100 | 100 | 100 | 100 |
6 | 5029 | 60.72 | 61.33 | 63.93 | 70.38 | 68.40 |
7 | 1330 | 62.09 | 72.22 | 70.12 | 72.33 | 90.46 |
8 | 3682 | 85.06 | 86.06 | 77.84 | 80.90 | 84.42 |
9 | 947 | 47.57 | 98.19 | 100 | 100 | 98.91 |
OA | 81.55 | 83.54 | 84.40 | 85.33 | 88.53 | |
AA | 75.75 | 82.15 | 82.80 | 83.92 | 89.12 | |
k×100 | 76.33 | 78.74 | 19.65 | 80.67 | 85.00 |
注: 粗体为最好的结果。
类别 | 数量 | GAN | GAN+BN | GAN+L2 | GAN+Dropout | GAN + HFAM |
---|---|---|---|---|---|---|
1 | 46 | 42.86 | 66.67 | 70.00 | 70.00 | 63.64 |
2 | 1428 | 75.35 | 74.83 | 74.82 | 77.57 | 92.72 |
3 | 830 | 80.92 | 86.36 | 84.67 | 82.52 | 88.81 |
4 | 237 | 57.58 | 49.25 | 50.00 | 57.14 | 70.15 |
5 | 483 | 80.85 | 89.77 | 92.86 | 95.24 | 95.60 |
6 | 730 | 94.56 | 97.18 | 55.56 | 94.48 | 95.97 |
7 | 28 | 50.00 | 50.00 | 100 | 62.50 | 71.43 |
8 | 478 | 93.88 | 98.94 | 98.93 | 100 | 100 |
9 | 20 | 20.00 | 50.00 | 50.00 | 100 | 100 |
10 | 972 | 69.58 | 78.85 | 74.17 | 75.55 | 82.73 |
11 | 2455 | 87.39 | 84.62 | 88.59 | 85.92 | 86.65 |
12 | 593 | 79.66 | 81.69 | 84.21 | 84.62 | 90.12 |
13 | 205 | 84.44 | 77.35 | 78.85 | 93.18 | 87.23 |
14 | 1265 | 96.81 | 99.52 | 99.54 | 99.54 | 99.58 |
15 | 386 | 54.23 | 57.50 | 58.12 | 61.78 | 84.44 |
16 | 93 | 60.00 | 75.00 | 89.47 | 85.71 | 77.27 |
OA | 79.92 | 82.43 | 83.22 | 84.01 | 89.61 | |
AA | 70.51 | 76.09 | 77.84 | 82.86 | 86.65 | |
k×100 | 77.26 | 80.02 | 80.91 | 81.80 | 88.14 |
注: 粗体为最好的结果。
图12 Indian Pines数据集准确率变化情况
Fig.12 Variation of the accuracy of Indian Pines dataset
图13 Pavia University数据集准确率变化情况
Fig.13 Variation of the accuracy of Pavia University dataset
图14 Pavia University数据集取0.1%标记样本不同方法的分类结果图
Fig.14 Classification maps provided by different approaches with 0.1% labeled samples on the Pavia University dataset
图15 Indian Pines数据集取1%标记样本不同方法的分类结果图
Fig.15 Classification maps provided by different approaches with 1% labeled samples on the Indian Pines dataset
本文通过分析深度学习模型产生过拟合的原因,利用对抗思想在过拟合和欠拟合之间寻找平衡点,解决了高光谱遥感图像在使用深度学习模型进行分类时,由于训练样本较少而导致分类准确率较低的问题。通过分析高光谱遥感图像的光谱特征和空间特征,设计了更加高效的网络结构。首先利用全连接网络提取邻域块中各个像元的光谱特征,然后用平均池化对空间特征进行融合,提高了特征提取的效率,减少了深度学习模型训练的时间。通过在两个不同数据集上进行的实验验证了本文所设计网络结构的高效性和鲁棒性,说明在设计深度学习网络结构时充分分析数据的特性是必要的。通过与其他过拟合抑制算法对比,验证了本文过拟合抑制算法的先进性,说明在网络模型拟合数据的过程中,利用对抗思想抑制拟合过程能够改善拟合效果,防止过度拟合。
深度学习模型通常包含大量的参数,能够拟合复杂的输入特征,但是当训练样本较少时,无约束地拟合训练样本会导致过拟合现象严重,即使训练样本较多的情况下,过拟合现象也普遍存在。通常抑制过拟合的方法是减小参数的值,认为小的参数值能够使模型更加平滑,对输入数据的拟合效果更好。然而,真实数据的分布情况复杂多变,该方法对过拟合抑制的作用存在很大的局限性。并且,产生过拟合的原因不仅是模型参数的值较大,模型参数的方差较大也是导致过拟合的原因之一,只减小参数的值难以达到理想的效果。此外,该类方法没有针对输入数据的特征来抑制过拟合,对分布较为复杂的数据难以有效发挥作用。本文通过对模型高维特征均值最小化,并与拟合过程不断对抗,不仅合理减小了模型参数的值和方差,还充分考虑了输入数据的真实分布情况。在两个标准数据集Indian Pines和Pavia University上进行的实验表明,本文算法比目前表现最好的过拟合抑制算法,准确率分别提高了5.60%和3.20%。
实验中发现本文算法在训练样本较多时对过拟合的抑制效果不够明显,并且在训练样本较少时,训练误差和测试误差相差还比较大。下一步工作将针对深度学习模型过拟合问题进行更加详细的理论研究,并在不同类型的数据集上进行实验,进一步提高高光谱遥感图像在少量训练样本情况下的分类准确率。
Arjovsky M, Chintala S and Bottou L. 2017. Wasserstein generative adversarial networks//Proceedings of the 34th International Conference on Machine Learning. Sydney, Australia: PMLR: 214-223 [百度学术]
Burden F and Winkler D. 2008. Bayesian regularization of neural networks/ (/Livingstone D J, ed. Artificial Neural Networks. [s.l.]: Humana Press: 23-42) [DOI: 10.1007/978-1-60327-101-1_3] [百度学术]
Burnham K P and Anderson D R. 2002. Model Selection and Multimodel Inference. New York: Springer [DOI: 10.1007/b97636] [百度学术]
Camps-Valls G, Gomez-Chova L, Muñoz-Marí J, Vila-Francés J and Calpe-Maravilla J. 2006. Composite kernels for hyperspectral image classification. IEEE Geoscience and Remote Sensing Letters, 3(1): 93-97 [DOI: 10.1109/LGRS.2005.857031] [百度学术]
Chen C, Chen N and Peng J T. 2016a. Nearest regularized joint sparse representation for hyperspectral image classification. IEEE Geoscience and Remote Sensing Letters, 13(3): 424-428 [DOI: 10.1109/LGRS.2016.2517095] [百度学术]
Chen C, Li W, Su H J and Liu K. 2014a. Spectral-spatial classification of hyperspectral image based on kernel extreme learning machine. Remote Sensing, 6(6): 5795-5814 [DOI: 10.3390/rs6065795] [百度学术]
Chen X, Duan Y, Houthooft R, Schulman J, Sutskever I and Abbeel P. 2016b. Infogan: interpretable representation learning by information maximizing generative adversarial nets//Proceedings of the 30th Conference on Neural Information Processing Systems. Barcelona, Spain: NIPS: 2172-2180 [百度学术]
Chen Y, Nasrabadi N M and Tran T D. 2013. Hyperspectral image classification via kernel sparse representation. IEEE Transactions on Geoscience and Remote Sensing, 51(1): 217-231 [DOI: 10.1109/TGRS.2012.2201730] [百度学术]
Chen Y, Zhu L, Ghamisi P, Jia X P, Li G Y and Tang L. 2017. Hyperspectral images classification with Gabor filtering and convolutional neural network. IEEE Geoscience and Remote Sensing Letters, 14(12): 2355-2359 [DOI: 10.1109/LGRS.2017.2764915] [百度学术]
Chen Y S, Jiang H L, Li C Y, Jia X P and Ghamisi P. 2016c. Deep feature extraction and classification of hyperspectral images based on convolutional neural networks. IEEE Transactions on Geoscience and Remote Sensing, 54(10): 6232-6251 [DOI: 10.1109/TGRS.2016.2584107] [百度学术]
Chen Y S, Lin Z H, Zhao X, Wang G and Gu Y F. 2014b. Deep learning-based classification of hyperspectral data. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 7(6): 2094-2107 [DOI: 10.1109/JSTARS.2014.2329330] [百度学术]
Cogswell M, Ahmed F, Girshick R B, Zitnick L and Batra D. 2016. Reducing overfitting in deep networks by decorrelating representations//Proceedings of the 4th International Conference on Learning Representations. San Juan: ICLR [百度学术]
Cui B G, Ma X D and Xie X Y. 2017. Hyperspectral image de-noising and classification with small training samples. Journal of Remote Sensing, 21(5): 728-738 [百度学术]
崔宾阁, 马秀丹, 谢小云. 2017. 小样本的高光谱图像降噪与分类. 遥感学报, 21(5): 728-738 [DOI: 10.11834/jrs.20176239] [百度学术]
Demarez V. 1999. Seasonal variation of leaf chlorophyll content of a temperate forest. Inversion of the PROSPECT model. International Journal of Remote Sensing, 20(5): 879-894 [DOI: 10.1080/014311699212975] [百度学术]
Du P J, Xia J S, Xue Z H, Tan K, Su H J and Bao R. 2016. Review of hyperspectral remote sensing image classification. Journal of Remote Sensing, 20(2): 236-256 [百度学术]
杜培军, 夏俊士, 薛朝辉, 谭琨, 苏红军, 鲍蕊. 2016. 高光谱遥感影像分类研究进展. 遥感学报, 20(2): 236-256 [DOI: 10.11834/jrs.20165022] [百度学术]
Giacinto G, Roli F and Fumera G. 2000. Design of effective multiple classifier systems by clustering of classifiers//Proceedings of the 15th International Conference on Pattern Recognition. Barcelona, Spain: IEEE: 160-163 [DOI: 10.1109/ICPR.2000.906039] [百度学术]
Goodfellow I J, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, Courville A and Bengio Y. 2014. Generative adversarial nets//Proceedings of the 27th International Conference on Neural Information Processing Systems. Montreal, Canada: NIPS: 2672-2680 [百度学术]
Hammer B. 2001. Neural smithing—supervised learning in feedforward artificial neural networks. Pattern Analysis and Applications, 4(1): 73-74 [DOI: 10.1007/s100440170029] [百度学术]
He N J, Paoletti M E, Haut J M, Fang L Y, Li S T, Plaza A and Plaza J. 2019. Feature extraction with multiscale covariance maps for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 57(2): 755-769 [DOI: 10.1109/TGRS.2018.2860464] [百度学术]
Ioffe S. 2017. Batch renormalization: towards reducing minibatch dependence in batch-normalized models//Proceedings of the 31st Conference on Neural Information Processing Systems. Long Beach, USA: NIPS: 1945-1953 [百度学术]
Ioffe S and Szegedy C. 2015. Batch normalization: accelerating deep network training by reducing internal covariate shift//Proceedings of the 32nd International Conference on Machine Learning. Lille, France: ICML: 448-456 [百度学术]
Krizhevsky A, Sutskever I and Hinton G E. 2012. Imagenet classification with deep convolutional neural networks //Proceedings of the 25th International Conference on Neural Information Processing Systems. Lake Tahoe, Nevada, USA: NIPS: 1097-1105 [百度学术]
Lever J, Krzywinski M and Altman N. 2016. Points of significance: model selection and overfitting. Nature Methods, 13(9): 703-704 [DOI: 10.1038/nmeth.3968] [百度学术]
Liu Z, Tang B, He X F, Qiu Q C and Liu F. 2017. Class-specific random forest with cross-correlation constraints for spectral-spatial hyperspectral image classification. IEEE Geoscience and Remote Sensing Letters, 14(2): 257-261 [DOI: 10.1109/LGRS.2016.2637561] [百度学术]
Mao X D, Li Q, Xie H R, Lau R Y K, Wang Z and Paul Smolley S. 2017. Least squares generative adversarial networks//Proceedings of 2017 IEEE International Conference on Computer Vision. Venice, Italy: IEEE: 2813-2821 [DOI: 10.1109/ICCV.2017.304] [百度学术]
Pan B, Shi Z W and Xu X. 2017. R-VCANet: a new deep-learning-based hyperspectral image classification method. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 10(5): 1975-1986 [DOI: 10.1109/JSTARS.2017.2655516] [百度学术]
Qureshi R, Uzair M, Khurshid K and Yan H. 2019. Hyperspectral document image processing: applications, challenges and future prospects. Pattern Recognition, 90: 12-22 [DOI: 10.1016/j.patcog.2019.01.026] [百度学术]
Salimans T, Goodfellow I, Zaremba W, Cheung V, Radford A and Chen X. 2016. Improved techniques for training GANs//Proceedings of the 30th Conference on Neural Information Processing Systems. Barcelona, Spain: NIPS: 2234-2242 [百度学术]
Soltani-Farani A, Rabiee H R and Hosseini S A. 2015. Spatial-aware dictionary learning for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 53(1): 527-541 [DOI: 10.1109/TGRS.2014.2325067] [百度学术]
Srivastava N, Hinton G, Krizhevsky A, Sutskever I and Salakhutdinov R. 2014. Dropout: a simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research, 15(1): 1929-1958 [百度学术]
Teke M, Deveci H S, Haliloğlu O, Gürbüz S Z and Sakarya U. 2013. A short survey of hyperspectral remote sensing applications in agriculture//Proceedings of the 2013 6th International Conference on Recent Advances in Space Technologies. Istanbul, Turkey: IEEE: 171-176 [DOI: 10.1109/RAST.2013.6581194] [百度学术]
Wu B, Zhu Y, Huang X and Li J Y. 2016. Generalization of spectral fidelity with flexible measures for the sparse representation classification of hyperspectral images. International Journal of Applied Earth Observation and Geoinformation, 52: 275-283 [DOI: 10.1016/j.jag.2016.06.006] [百度学术]
Yuen P W and Richardson M. 2010. An introduction to hyperspectral imaging and its application for security, surveillance and target acquisition. The Imaging Science Journal, 58(5): 241-253 [DOI: 10.1179/174313110x12771950995716] [百度学术]
Zhan Y, Hu D, Wang Y T and Yu X C. 2018. Semisupervised hyperspectral image classification based on generative adversarial networks. IEEE Geoscience and Remote Sensing Letters, 15(2): 212-216 [DOI: 10.1109/LGRS.2017.2780890] [百度学术]
Zhang K, Hei B Q, Zhou Z and Li S Y. 2018. CNN with coefficient of variation-based dimensionality reduction for hyperspectral remote sensing images classification. Journal of Remote Sensing, 22(1): 87-96 [百度学术]
张康, 黑保琴, 周壮, 李盛阳. 2018. 变异系数降维的CNN高光谱遥感图像分类. 遥感学报, 22(1): 87-96 [DOI: 10.11834/jrs.20187075] [百度学术]
Zhao W Z and Du S H. 2016. Learning multiscale and deep representations for classifying remotely sensed imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 113: 155-165 [DOI: 10.1016/j.isprsjprs.2016.01.004] [百度学术]
Zhong Z L and Li J. 2018. Generative adversarial networks and probabilistic graph models for hyperspectral image classification//Proceedings of the 32nd AAAI Conference on Artificial Intelligence. New Orleans: AAAI: 8191-8193 [百度学术]
Zhong Z L, Li J, Luo Z M and Chapman M. 2018. Spectral-spatial residual network for hyperspectral image classification: a 3-D deep learning framework. IEEE Transactions on Geoscience and Remote Sensing, 56(2): 847-858 [DOI: 10.1109/TGRS.2017.2755542] [百度学术]
Zhu L, Chen Y S, Ghamisi P and Benediktsson J A. 2018. Generative adversarial networks for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 56(9): 5046-5063 [DOI: 10.1109/TGRS.2018.2805286] [百度学术]
Zur R M, Jiang Y L, Pesce L L and Drukker K. 2009. Noise injection for training artificial neural networks: a comparison with weight decay and early stopping. Medical Physics, 36(10): 4810-4818 [DOI: 10.1118/1.3213517] [百度学术]
相关作者
相关机构