首页 >  , Vol. , Issue () : -

摘要

全文摘要次数: 400 全文下载次数: 485
引用本文:

DOI:

10.11834/jrs.20219192

收稿日期:

2019-06-09

修改日期:

2019-10-16

PDF Free   EndNote   BibTeX
具备过拟合抑制的生成式对抗网络模型构建及遥感分类应用
张健, 保文星
北方民族大学
摘要:

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%,相比于其它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%。

Construction of Generative Adversarial Network Model with Overfitting Suppression and Application of Remote Sensing Classification
Abstract:

Objective 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. When the training sample is small, there will be overfitting. The accuracy of the model on the test set is lower than the accuracy on the training set. In order to suppress overfitting, the researchers proposed overfitting suppression methods such as weight decay and dropout. However, these methods need to work in a specific environment, and the suppression effect on overfitting is limited. Based on this, this paper proposes an overfitting suppression algorithm based on the Generative Ad-versarial Networks to suppress the overfitting phenomenon of the model. Method First, construct a spatial neighborhood block for the standard dataset and split it into labeled samples, unlabeled samples and test samples, and then send the labeled samples and unlabeled samples to the Gen-erative Adversarial Networks for training. When input, the pixels in the neighborhood block are independent-ly 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 connected to the output layer to obtain the classification result. Since the overfitting is caused by the large value and variance of the network parameters, the large parameter values enable the model to fit more samples. Therefore, in each iteration, the network is first fitted to the data by labeled samples, then use the optimizer to minimize the mean of the high-dimensional features. This process will re-update the network parameters, reduce the value and vari-ance of the parameters, and thus suppress the overfitting. Result The algorithm was applied to two standard datasets, Indian Pines and Pavia University. The 1% labeled samples were randomly selected for training. The overall classification accuracy rate was 89.61% and 98.79%, respectively, which was better than several algorithms with better performance. The Indian Pines dataset randomly selected 1% labeled samples, Pavia University dataset randomly selected 0.1% labeled samples, compared with several commonly used overfitting suppression methods Batch Normalization(BN), L2 regularization and Dropout, the overfitting suppression algorithm in this paper is 5.60% and 3.20% higher than the best Dropout, respectively. Conclusion The Generative Adversarial Networks model designed for the characteristics of hyperspectral data can make good use of the spectral and spatial features of hyperspectral images. The proposed overfitting sup-pression algorithm can significantly improve the classification performance of the model. However, when the number of labeled samples is large, the overfitting suppression effect of the algorithm is not obvious, and further research is needed.

本文暂时没有被引用!

欢迎关注学报微信

遥感学报交流群 分享按钮