在高光谱遥感图像分类中,需要大量的训练样本对分类器进行训练,然而对样本标记非常困难并且耗时、昂贵。针对样本标记困难的问题,提出了自适应的样本不确定性与代表性相结合的主动学习选择训练样本。样本的不确定性是利用最优标号与次...在高光谱遥感图像分类中,需要大量的训练样本对分类器进行训练,然而对样本标记非常困难并且耗时、昂贵。针对样本标记困难的问题,提出了自适应的样本不确定性与代表性相结合的主动学习选择训练样本。样本的不确定性是利用最优标号与次优标号(best vs second-best,BvSB)的方法计算。用期望最大(expectation maximization,EM)聚类计算样本的代表性。然后将样本的不确定性与代表性通过自适应权重相结合,从而选出含信息量最大的未标注样本加入进行人工标注,并加入到训练样本。通过实验表明,此方法性能更加稳定,准确率也有一定的提高。展开更多
Two revised regional importance measures(RIMs),that is,revised contribution to variance of sample mean(RCVSM)and revised contribution to variance of sample variance(RCVSV),are defined herein by using the revised means...Two revised regional importance measures(RIMs),that is,revised contribution to variance of sample mean(RCVSM)and revised contribution to variance of sample variance(RCVSV),are defined herein by using the revised means of sample mean and sample variance,which vary with the reduced range of the epistemic parameter.The RCVSM and RCVSV can be computed by the same set of samples,thus no extra computational cost is introduced with respect to the computations of CVSM and CVSV.From the plots of RCVSM and RCVSV,accurate quantitative information on variance reductions of sample mean and sample variance can be read because of reduced upper bound of the range of the epistemic parameter.For general form of quadratic polynomial output,the analytical solutions of the original and the revised RIMs are given.Numerical example is employed and results demonstrate that the analytical results are consistent and accurate.An engineering example is applied to testify the validity and rationality of the revised RIMs,which can give instructions to the engineers about how to reduce variance of sample mean and sample variance by reducing the range of epistemic parameters.展开更多
文摘在高光谱遥感图像分类中,需要大量的训练样本对分类器进行训练,然而对样本标记非常困难并且耗时、昂贵。针对样本标记困难的问题,提出了自适应的样本不确定性与代表性相结合的主动学习选择训练样本。样本的不确定性是利用最优标号与次优标号(best vs second-best,BvSB)的方法计算。用期望最大(expectation maximization,EM)聚类计算样本的代表性。然后将样本的不确定性与代表性通过自适应权重相结合,从而选出含信息量最大的未标注样本加入进行人工标注,并加入到训练样本。通过实验表明,此方法性能更加稳定,准确率也有一定的提高。
基金supported by the National Natural Science Foundation of China(Grant No.51175425)the Special Research Fund for the Doctoral Program of Higher Education of China(Grant No.20116102110003)
文摘Two revised regional importance measures(RIMs),that is,revised contribution to variance of sample mean(RCVSM)and revised contribution to variance of sample variance(RCVSV),are defined herein by using the revised means of sample mean and sample variance,which vary with the reduced range of the epistemic parameter.The RCVSM and RCVSV can be computed by the same set of samples,thus no extra computational cost is introduced with respect to the computations of CVSM and CVSV.From the plots of RCVSM and RCVSV,accurate quantitative information on variance reductions of sample mean and sample variance can be read because of reduced upper bound of the range of the epistemic parameter.For general form of quadratic polynomial output,the analytical solutions of the original and the revised RIMs are given.Numerical example is employed and results demonstrate that the analytical results are consistent and accurate.An engineering example is applied to testify the validity and rationality of the revised RIMs,which can give instructions to the engineers about how to reduce variance of sample mean and sample variance by reducing the range of epistemic parameters.
