基于不确定性的大样本集抽样设计

本文研究是针对全球森林覆盖变化检测系统的大训练样本选择的优化问题。样本自动标记系统自动标记了"千万个"代表森林和非森林的像元。为了提高精度和效率,我们需要从千万个样本的大样本集中选择对训练分类器有益的样本。本文提出了两种方法:第一种方法是传统的分层等距离采样;第二种方法是基于不确定性的样本优化,通过研究相同样本间的相似性度量指标、样本不确定性度量指标及样本特征空间分布,来实现基于不确定性的训练样本选择策略。实验通过精度评价验证了基于不确定性的采样策略能比传统的分层等距离采样策略获得好的效果。

Sampling Design for the Large Sample Set Based on Uncertainty

The research aims at solving the problem about optimizing large training sample set in global vegetation-classification system. Tens of millions of samples representing forests and non-forests are labeled by automatically training samples procedure. In order to improve the precision and efficiency, key samples should be picked up and this has a large contribution to prediction. Two methods have been put forward in this paper: stratified system sampling strategy and uncertainty-based sample set optimization. The latter one selects training samples based on uncertainty by examining the uncertainty measure of samples and the spatial distribution of their features. Through precision evaluation, the experiments validated that the uncertainty-based sampling strategy can achieve better results than the stratified system sampling strategy.