主题模型建模框架下的高分辨率遥感影像半监督分类研究

Automatic classification of very high resolution (VHR) remote sensing images provides an effective means for a great variety of applications such as urban planning, environmental monitoring and disaster information extraction. However, the efficiency of existing object-based classification methods is limited due to the complex process of multi-scale segmentation. Another kind of methods based on spatial analysis of neighborhood suffers from the limitation of low self-adaptive ability. Learning from successful experiences of topic models in text analysis, this project will apply topic models to classify VHR remote sensing images, in order to overcome the shortcomings of the traditional methods. Considering the facts that basic topic models can only model discrete visual features and use the supervised information attached to document to guide learning, we will firstly establish a novel classification framework based on topic models, which is able to directly model multidimensional continuous visual features. Then, in the proposed framework, two semi-supervised approaches will be further developed to study the classification problems under the condition that partially labeled pixels are available. The main research content of this project includes: (1) establishing a classification framework that can model multidimensional continuous visual features using Gaussian-latent Dirichlet allocation model; (2) proposing a method to perform joint max-margin learning and maximum likelihood estimation for semi-supervised classification of VHR multispectral remote sensing images; (3) presenting a hybrid generative-discriminative approach for semi-supervised classification of VHR multispectral remote sensing images. Research results have great significance to improving classification accuracy and promoting the use of VHR remote sensing images in disaster monitoring, and will further deepen the research and application of topic models and semi-supervised learning.

高分辨率遥感影像自动分类为开展城市规划、环境变化检测以及灾害信息提取等提供了有效的技术手段。然而，现有面向对象分类方法多尺度分割难、基于邻域空间分析的方法自适应能力弱。本项目借鉴主题模型在文本分析领域中的成功范例，将其应用到高分辨率遥感影像分类，以克服传统方法的不足。针对该类模型只能够建模离散视觉特征和文档层面监督信息的瓶颈，构建面向多维连续视觉特征建模的主题模型框架，并在此基础上研究两种高分辨率遥感影像半监督分类算法，探索部分标注像素样本给定下的分类问题。研究内容包括：(1) 构建针对多维连续视觉特征建模的高斯-潜狄利克雷分配模型；(2) 类别边界与样本似然联合极大化的遥感影像半监督分类方法；(3) 综合判别式和生成式模型的遥感影像半监督分类方法。研究成果对于提高高分辨率遥感影像分类精度、推进其在灾害监测等方面的应用具有重要的意义，并将进一步拓展主题模型和半监督学习的研究和应用。

高分辨率遥感影像自动分类为开展城市规划、环境变化检测以及灾害信息提取等提供了有效的技术手段。然而，现有面向对象分类方法多尺度分割难、基于邻域空间分析的方法自适应能力弱。本项目借鉴主题模型在文本分析领域和自然图片处理中的成功范例，将其应用到高分辨率遥感影像分类，以克服传统方法的不足。研究内容包括：(1) 针对传统主题模型只能够建模离散视觉特征，建立了面向多维连续视觉特征建模的高斯-潜狄利克雷分配模型多光谱遥感影像分类框架；(2) 在多尺度分割面向对象分类框架下，借助于潜狄利克雷分配模型分析手段，提出了一种自适应的非监督分类方法；(3) 在生产式主题模型和判别式SVM联合建模框架下，提出一种融合多尺度光谱－空间－语义特征的高分辨率遥感影像分类方法；(4) 针对传统主题模型只能够建模文档层面监督信息的瓶颈，建立了嵌入像元监督信息的最大熵判别式潜狄利克雷分配模型半监督分类方法，将其应用于多光谱和高光谱遥感影像分类；(5) 联合主题模型和自学习过程，提出了一种高分辨率遥感影像尺度自适应半监督分类方法。研究成果对于提高高分辨率遥感影像分类精度、推进其在灾害监测等方面的应用具有重要的意义，并将进一步拓展主题模型和面向对象分类的研究和应用。