基于集成学习的PolSAR标签噪声研究

发布时间：2018-04-12 16:16

  本文选题：PolSAR + 集成学习　； 参考：《西安电子科技大学》2014年硕士论文

【摘要】：极化合成孔径雷达(Polarimetric Synthetic Aperture Radar,PolSAR)是一种多参数、多通道的成像雷达系统,因其全天时,全天候,高分辨的优势而得到广泛的应用。基于机器学习的PolSAR图像分类方法取得了很高的分类精度,但是当有标签噪声存在时,分类结果会受到很大的影响。本文基于集成学习,针对PolSAR图像分类中标签噪声的问题,进行了深入的研究,主要包括以下三方面的内容:1.结合PolSAR图像的偏振参数、散射、纹理特征,提出了一种基于AdaBoost的PolSAR图像监督分类算法(Knn.Ada Boost)。此方法利用PolSAR图像的偏振参数、极化散射特征和图像的纹理特征,作为Ada Boost的输入特征,Knn.Ada Boost算法预先通过K nn计算PolSAR图像中每个像素的抗噪因子,根据抗噪因子修改Ada Boost算法中的样本权值更新策略。实验采用了一组模拟PolSAR数据和五组真实PolSAR数据,实验结果表明,K nn.Ada Boost算法提高了AdaBoost的分类精度,具有很好的抗噪性能。2.在Knn.Ada Boost的工作基础上,提出了一种基于Ada Boost的PolSAR图像半监督分类算法(Semi.Knn.AdaBoost)。在Knn.Ada Boost的框架下,引入Wishart距离度量,在每一次迭代结束时,根据有标记样本计算获得Wishart聚类中心,从预测标记中选择距离Wishart聚类中心最近的若干个样本,分别加入对应的类别进入下一次迭代。实验采用一组模拟PolSAR数据和五组真实PolSAR数据,结果表明,Semi.Knn.Ada Boost丰富了训练样本,分类正确率有一定的提升。3.在PolSAR图像分类问题中,提出了一种基于集成学习的标签噪声水平预测方法EEL(Estimated by Ensemble Learning)。采用PolSAR图像的相干矩阵中九个元素作为特征,利用不同的分类算法,学习得到相互独立的分类器,用这些分类器分别对标记样本分类,然后用多数投票和全投票的策略判定一个已标记样本是否是噪声,多数投票策略即对一个样本的预测,如果超过半数分类器的分类结果是相同的,则认为这个已标记样本不是噪声,否则是噪声;全投票策略只认定所有分类器投票结果相同时,此样本才不是噪声,否则是噪声。实验采用三组UCI数据和四组模拟的PolSAR数据,结果表明,在标签噪声水平比较低时,此方法能够正确的预测,而标签噪声水平比较高时,预测出的标签噪声水平则不是很准确。本文工作得到了国家自然科学基金(No.61173092)、新世纪优秀人才支持计划(No.66ZY110)和陕西省科学技术研究发展计划项目(No.2013KJXX-64)资助。
[Abstract]:Polarimetric Synthetic Aperture Radarr (PolSAR) is a multi-parameter, multi-channel imaging radar system, which is widely used because of its advantages of all-weather, all-weather and high-resolution.The PolSAR image classification method based on machine learning has achieved high classification accuracy, but when there is label noise, the classification results will be greatly affected.Based on ensemble learning, this paper focuses on the problem of label noise in PolSAR image classification, including the following three aspects: 1.Based on the polarization parameters, scattering and texture features of PolSAR images, a supervised classification algorithm for PolSAR images based on AdaBoost is proposed.Using polarization parameters, polarization scattering features and texture features of PolSAR images, the Knn.Ada Boost algorithm is used as the input feature of Ada Boost to calculate the anti-noise factor of each pixel in PolSAR image.The sample weight updating strategy in Ada Boost algorithm is modified according to the anti-noise factor.A set of simulated PolSAR data and five groups of real PolSAR data are used in the experiment. The experimental results show that the K nn.Ada Boost algorithm improves the classification accuracy of AdaBoost and has a good anti-noise performance.Based on the work of Knn.Ada Boost, a semi-supervised PolSAR image classification algorithm based on Ada Boost is proposed.In the framework of Knn.Ada Boost, the Wishart distance metric is introduced. At the end of each iteration, the Wishart cluster center is obtained according to the calculation of labeled samples, and several samples closest to the Wishart cluster center are selected from the prediction markers.Add corresponding categories to the next iteration.A set of simulated PolSAR data and five groups of real PolSAR data are used in the experiment. The results show that Semi.Knn.Ada Boost enriches the training samples and improves the classification accuracy. 3.In the problem of PolSAR image classification, an ensemble learning based label noise prediction method, EEL(Estimated by Ensemble learning, is proposed.Using nine elements in the coherent matrix of PolSAR image as features and using different classification algorithms, independent classifiers are obtained, and these classifiers are used to classify the labeled samples respectively.Then the majority voting strategy is used to determine whether a marked sample is noisy or not, and the majority voting strategy is the prediction of a sample, if the classification results of more than half of the classifiers are the same.It is considered that the labeled sample is not noise, otherwise it is noise; if the voting result of all classifiers is the same, the sample is not noise, otherwise it is noise.Three groups of UCI data and four groups of simulated PolSAR data are used in the experiment. The results show that this method can correctly predict the label noise level when the label noise level is low, but the predicted label noise level is not very accurate when the label noise level is high.This work is supported by the National Natural Science Foundation No. 61173092, the New Century Talent support Program No. 66ZY110) and the Shaanxi Provincial Science and Technology Research and Development Program Project No. 2013KJXX-64).
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN957.52

【相似文献】

相关期刊论文 前3条

1 代大海;王雪松;肖顺平;李永祯;;PolSAR有源假目标干扰的鉴别与对消[J];电子学报;2007年09期

2 李洪忠;陈劲松;王超;张红;;基于相似性的POLSAR占优散射归类及非监督聚类[J];电子与信息学报;2012年06期

3 ;[J];;年期

相关硕士学位论文 前2条

1 陈国栋;基于集成学习的PolSAR标签噪声研究[D];西安电子科技大学;2014年

2 杨丹;POLSAR原始回波信号仿真研究[D];哈尔滨工业大学;2007年

，

本文编号：1740480