【计算机科学】【2015】有限标记训练数据的遥感图像自动判读

本文为加拿大滑铁卢大学（作者：Fan Li）的博士论文，共134页。

遥感图像自动判读已经研究了十多年。在早期，大多数工作都是基于这样的假设：有足够的标记样本用于训练。然而，地面真实情况的收集是一项非常繁琐和耗时的工作，有时也非常昂贵，特别是在通常依靠野外调查来收集地面真实情况的遥感领域。近年来，随着先进机器学习技术的发展，有限地面真实情况的遥感图像解译引起了遥感和计算机科学领域研究者的关注。针对有限地面真实情况问题，提出了三种不同的解释方法，即特征提取、分类和分割。首先，研究了遥感图像分类中常用的预处理方法——特征提取技术。提出了一种基于集成局部流形学习的联合特征提取与分类框架。其次，研究了有限标记训练数据情况下的分类器，提出了一种性能优于现有分类方法的增强集成学习方法。第三，研究了基于未标记样本和空间信息的图像分割技术。提出了一种半监督自训练方法，该方法能够根据训练样本的个数进行扩展，从而迭代地提高分类性能。实验表明，该方法在分类精度方面优于现有的基准遥感数据集分类方法。

Automated remote sensing imageinterpretation has been investigated for more than a decade. In early years,most work was based on the assumption that there are sufficient labeled samplesto be used for training. However, ground-truth collection is a very tedious andtime-consuming task and sometimes very expensive, especially in the field ofremote sensing that usually relies on field surveys to collect ground truth. Inrecent years, as the development of advanced machine learning techniques,remote sensing image interpretation with limited ground-truth has caught theattention of researchers in the fields of both remote sensing and computerscience. Three approaches that focus on different aspects of the interpretationprocess, i.e., feature extraction, classification, and segmentation, areproposed to deal with the limited ground truth problem. First, featureextraction techniques, which usually serve as a preprocessing step for remotesensing image classification are explored. Instead of only focusing on featureextraction, a joint feature extraction and classification framework is proposedbased on ensemble local manifold learning. Second, classifiers in the case oflimited labeled training data are investigated, and an enhanced ensemblelearning method that outperforms state-of-the-art classification methods isproposed. Third, image segmentation techniques are investigated, with the aidof unlabeled samples and spatial information. A semi-supervised self-trainingmethod is proposed, which is capable of expanding the number of trainingsamples by its own and hence improving classification performance iteratively.Experiments show that the proposed approaches outperform stateof-the-arttechniques in terms of classification accuracy on benchmark remote sensingdatasets.

1. 引言
2. 文献回顾
3. 数据集
4. 集成局部流形学习的联合特征提取与分类

更多精彩文章请关注公众号：