联合空谱上下文的高光谱遥感图像低秩表示分类理论与算法

项目名称： 联合空谱上下文的高光谱遥感图像低秩表示分类理论与算法

项目编号： No.61471199

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 无线电电子学、电信技术

项目作者： 吴泽彬

作者单位： 南京理工大学

项目金额： 83万元

中文摘要： 高光谱图像分类是遥感信息处理领域的研究热点。高光谱图像维数高，波段间相关性大，空间分辨率低，类内差异性强，存在噪声等退化情况，且训练样本相对不足，分类精度很难保证。现有的稀疏表示分类方法主要利用结构化稀疏先验来设计分类模型,只考虑了高光谱像元局部邻域内的部分相关性，缺乏对数据全局特性的建模，其精确性有待提高。本项目研究空谱上下文联合的高光谱低秩表示分类理论和算法，主要内容包括：高光谱图像的鉴别性结构化低秩表示字典学习，空谱上下文联合的高光谱数据图结构化稀疏性建模，利用数据的全局相关性和内在局部结构建立空谱联合的高光谱低秩表示分类模型，基于凸分析理论设计快速分类算法，结合CPU/GPU异构混合执行模式进行多级协同并行优化，并面向地质勘探和水域环境监测进行应用试验和测试验证。为提高高光谱遥感信息处理和定量解译的精度奠定基础，推动高光谱遥感在对地观测、深空探测领域的实际应用，具有广阔的应用前景。

中文关键词： 高光谱图像；分类；低秩表示；空谱上下文；稀疏性

英文摘要： Hyperspectral image classification is a hot topic in the field of remote sensing information processing. Because of the image characteristics of high dimensionality, strong spectral correlation, big inter-class difference and insufficience of training samples, it is difficult to realize high precision hyperspectral classification. Existing sparse representation classification methods mainly use structured sparsity to design hyperspectral classification model, which take into consideration the partial local correlation without global structure. In order to further improve the classification accuracy, discriminative structured dictionary learning method for low rank representation of hyperspectral image is studied, graph structured sparsity with spectral-spatial joint context is expoited, low rank representation classification model is advanced based on the global correlation and intrinsic local structure of hyperspectral image, fast algorithm and its parallel optimization for hyperspectral classification is designed on CPU/GPU heterogeneous execution mode. A novel hyperspectral image classification method based on low-rank representation and spatial-spectral context is proposed, which will lay the foundations for efficient hyperspectral information processing and quantitative interpretation, and have wide application perspectives in the fields of earth observation and deep space exploration.

英文关键词： Hyperspectral Image;Classification;Low Rank Representation;Spatial Spectral Context;Sparsity