高光谱遥感图像谱窗选择及联合分类研究

Hyperspectral remote sensing can be widely applied to key areas of national economy, such as crop yield estimation, geological survey, environmental monitoring etc., and is one of the important divisions of the China high-tech development plan. By considering the characteristics regarding the hyperspectral imagery, such as high redundancy, high dimensionality, etc., we propose to study effective methods and theories to increase the accuracy of hyperspectral image classification，through optimizing feature extraction and improving classifier design. The research will be focused on the spectral window modeling, the automatic spectral window searching, and the joint classification of hyperspectral multiple feature sets. Taking multi-category vegetation and road classification as research background, first we will build a model to describe the proposed spectral window by defining a set of relevant concepts. Next, we will investigate the quantitative relationship between the spectral window and the classification accuracy, forming a set of reasonable, complete, and well-defined mathematical presentation for the spectral window. Based on the defined spectral window model, then we will develop several window searching algorithms, and implement an algorithm via the differentiable mutual information. Using the AVIRIS and HYDICE hyperspectral data sets, we will justify the suitability and effectiveness of the proposed spectral window selection algorithms. Finally, we will modify the traditional Bayesian rule of classifiers combination to better fuse two hyperspectral feature sets. One of these feature set is extracted from the 'best' selected spectral window by a 'discriminative' criterion, and the other feather set is extracted by inversion of the hyperspectral sensing data with a 'generative' property. Joint using the 'discriminative' feature set and the 'generative' feature set has a potential to capture more information embedded in the remotely sensed data, and then lead to improved classification accuracy.

高光谱遥感可以广泛应用于农业估产、地质调查、环境监测等国民经济关键领域，是国家高科技发展的重要内容。本项目将紧密围绕高光谱图像的数据高冗余性，单像素点高维度等特点，从优化特征提取和改进分类器设计出发，研究提升高光谱遥感图像分类准确性的方法与理论。着重研究高光谱特征谱窗建模、谱窗自动搜寻算法、高光谱多特征集联合分类方法等。以自然植被和人工道路等地物分类为研究背景，定义高光谱特征谱窗概念，建立高光谱谱窗模型，揭示谱窗与分类准确率之间的定量关系，形成合理、完整与清晰的谱窗特征选择表达式。在此基础上，开发谱窗搜寻方法，实现基于可微互信息的自动特征谱窗搜索算法，利用AVIRIS和HYDICE等高光谱数据集，验证特征谱窗选择的合理性和有效性。最后，从特征谱窗数据中提取具有"可分性"特性的特征集，应用改进的贝叶斯分类器组合方法，与具有"生成"特性的高光谱反演特征集进行联合分类，达到提升分类准确性的目的。

高光谱遥感可以广泛应用于农业估产、地质调查、环境监测等国民经济关键领域，是国家高科技发展的重要内容。本项目紧密围绕高光谱图像的数据高冗余性，单像素点高维度等特点，从优化特征提取和改进分类器设计出发，研究了高光谱遥感图像分类方法与理论。重点研究了高光谱特征谱窗建模、谱窗自动搜寻算法、高光谱多特征集联合分类方法等。定义了高光谱特征谱窗概念，建立了高光谱谱窗模型，揭示出谱窗与分类准确率之间的定量关系。在此基础上，开发了一个谱窗搜寻方法，实现了一种基于可微互信息的自动特征谱窗搜索算法，利用AVIRIS和HYDICE等高光谱数据集，验证了特征谱窗选择的合理性和有效性。最后，从特征谱窗数据中提取具有“可分性”特性的特征集，应用改进的贝叶斯分类器组合方法，与具有“生成”特性的高光谱反演特征集进行了联合分类，达到了提升分类准确性的目的。