基于TM影像的单类分类算法比较研究

Comparing of One-Class Classifier Based on Remote Sensing Image

单类分类是指在只有一种类别样本的情况下,只通过这一类样本训练分类器,再用训练出的分类器对未知类别的样本进行类别判断.在遥感影像分类问题中,当某种类别样本无法获取或相对于其他样本数量很少时,就会产生各类样本数量不均衡的现象,传统的两类或多类分类方法将不能很好地适用;当只需要从图像中提取某种特定的类别时,标定大量非此类别的样本将花费不必要的时间,这时就需要用单类分类器来解决问题.因此,研究单类分类器在遥感影像分类问题中的应用有着重大的意义.该文集中讨论几种典型的单类分类算法,将他们应用于TM遥感影像进行比较分析实验.证明基于支撑域的单类分类方法OCSVM(one-class SVM)和基于密度的方法GDD(Gaussian Domain Descriptor)均可以完成针对遥感影像的地物分类,而BSVM方法因考虑了更多样本的信息,能够得到更好的分类结果.得出结论,在单类分类问题中,加入未知类别样本的信息,可以提高分类效果.而这些方法的缺点是参数多且分类结果对参数敏感,这些问题有待在今后进一步研究.

One-class classification means training the classifier on data of only one class,and then label the data to be classified with the classifier.In remote-sensing classification,there are situations when we can hardly or even not get samples of some classes,while the traditional classification methods will not be well adapted to the unbalanced number of samples.And for some applications,we may only be interested in a spe⁃cific class without considering other land types.Then labeling samples of all classes occur in the image may increase the classification difficulty and cost for labeling training data.These problems can be referred to as one-class classification.In this article,we introduce and analysis some typical approaches of one-class clas⁃sification,and then perform some experiment on TM remote sensed data.Experimental results show the one-class SVM(OCSVM)method and GDD method provide relatively better performance in remote sensing image classification,and the biased SVM(BSVM)method could get better results since combining both la⁃beled and unlabeled data for classifier training.Experimental results also indicate that unlabeled samples al⁃so provide useful information for the construction of classifiers.The disadvantage of these methods is that there are many parameters to be selected and the classification results are sensitive to parameters.These problems need to be further studied in the future.