基于异质多视图主动学习的高光谱地物分类

HYPERSPECTRAL IMAGE CLASSIFICATION BASED ON HETEROGENEOUS MULTI-VIEW ACTIVE LEARNING

高光谱遥感影像具有多源异质的属性特征,也面临着训练样本少、标记代价大的困难。拟提取空间形状、纹理等多种属性特征来构建多视图,开展基于异质多视图主动学习的高光谱地物分类研究。主要解决两个问题:1)提出一种新的基于多视图后验概率差异最小(MPPD)的样本查询策略。每个视图根据多元逻辑回归分类器预测样本的类别条件概率;根据全概率公式计算多视图下每个样本的后验概率;挑选后验概率差异最小的样本作为信息含量最大的样本。2)提出一种基于空间多尺度形状结构、以及纹理特征的异质多视图的构建方式。实验结果表明,提出的算法能够加快学习函数的收敛速度,以少量的信息含量大的标记样本来提高学习器的预测性能。

Hyperspectral remote sensing images have multi-source,heterogeneous characteristics,but face the problems of less samples and labeling difficulty. This paper intent to extract multiple types of attribute features,including spatial shape and texture,etc.,to construct multi-view and study the heterogeneous multi-view based active learning for hyperspectral image classification. Two main issues were included: 1) a new query strategy based on the minimum posteriori probability difference( MPPD) for multi-view active learning was proposed. Each view was used to predict the conditional probability of each sample according to the multinomial logistical regression classifier; the posterior probability of each sample under the multi-view was calculated according to the full probability formula; the samples with the minimum difference in posterior probability were selected as the most informative ones. 2) A heterogeneous view generation strategy was proposed based on the multi-scale spatial shape and texture features. The experimental results showed that the proposed algorithm could speed up the convergence of learning functions and improved the predictive performance of learners with a small number of labeled samples with large information content.