摘要
为了解决局部支持向量机算法KNNSVM存在的分类时间过长不利于具有海量数据量的高分辨率遥感图像分类的不足,提高KNNSVM的算法表现,提出了改进的基于不确定性的BKNNSVM算法.该算法利用二项式分布的共轭先验分布Beta分布根据近邻的分布情况推导该未标记样本属于正类或负类的概率大小,从而计算每一个未标记样本在类属性上的不确定性大小.再通过设置不确定性阈值的大小,对不确定性低于阈值的未标记样本直接采用KNN进行分类,而对高于阈值的样本利用其近邻建立局部支持向量机分类器进行分类.对高分辨率图像分类的实验结果表明:合适的阈值能够有效降低原始KNNSVM算法的时间开销,同时能保持KNNSVM分类精度高的特点.
In order to solve the deficiency about the algorithm of the local support vector machine ( KNNSVM ) which is very time-consuming on classifying the high resolution remote sensing images with mass data and improve the performance of the KNNSVM, the BKNNSVM algorithm based on uncertainty is proposed. The algorithm applies the binomial distribution of conjugate prior Beta distribution to estimate probability belong to the class or negative of each unlabeled sample through its nearest neighbor distribution. Then, with threshold, some unlabeled samples are classified by the KNN when their uncertainty value is less than threshold and others are classified by the local SVM when their uncertainty value is more than threshold. The experiments on the actual high resolution remote sensing images have shown that BNNSVM can decrease the time consuming effectively and keep the precision of the original KNNSVM with suitable threshold of uncertainty.
出处
《中南民族大学学报(自然科学版)》
CAS
北大核心
2016年第1期95-102,共8页
Journal of South-Central University for Nationalities:Natural Science Edition
基金
湖北省自然科学基金资助项目(PBZY14019)