基于样本类别确定度的半监督分类

在对遥感图像进行分类时，全监督算法往往需要足够的标记样本进行训练，然而标记的过程是耗时和昂贵的，相反收集大量的无标记样本是很容易的。为了在学习过程中能够有效利用未标记样本的信息，本文提出了基于样本类别确定度（CCS）的半监督分类算法。首先，利用多分类支持向量机（SVM）得到未标记样本属于各类别的确定度，有效地衡量了未标记样本类别可靠性；其次，对样本类别确定度进行预处理，提升利用未标记样本的安全性；最后，基于样本类别确定度设计了半监督线性判别分析（LDA）降维算法并对其进行核化，使得样本在降维后的子空间更具有可分性，并根据降维后的数据特点，采用最近邻分类器对新样本进行分类。利用真实的合成孔径雷达（SAR）图像进行测试，验证了在标记样本较少的情况下，本文算法在性能上优于全监督和其他半监督算法，并能够快速收敛。

Semi-supervised classification based on class certainty of samples

The performance of supervised learning based algorithms can decrease dramatically in the classification of remote sensing images if labeled samples are insufficient. The collection of labeled samples is generally time-consuming and expensive, though unlabeled samples can be relatively easily obtained. To utilize the information of unlabeled samples in the learning process, this paper proposes a novel semi-supervised classification algorithm based on class certainty of samples (CCS). First, a multi-class support vector machine (SVM) is employed to determine the class certainty of unlabeled samples, which effectively measure the class reliability of unlabeled samples. Then, the pre-processing of sample classification is carried out to enhance the security of unlabeled samples. Finally, a new semi-supervised linear discriminant analysis (LDA) is proposed based on the sample class certainty and results in improved separability of the samples in the projection subspace. Moreover, the semi-supervised LDA can be extended to nonlinear dimensional reduction by combining the class certainty and the kernel based methods. For classification of the testing samples, the nearest neighbor classifier is adopted. In order to assess the effectiveness of the proposed algorithm, several experiments are carried out on the actual synthetic aperture radar (SAR) images in comparison with other supervised and semi-supervised algorithms. Using real SAR images, it is proved that the proposed algorithm is superior to all supervised and other semi supervised algorithms in the case of less marked samples. And it can converge quickly.