顾及异方差与空间约束的高光谱混合像元模糊聚类分解方法研究

Hyperspectral unmixing is a popular yet challenging research topic, with the key issue being efficient model construction and optimization to obtain accurate model parameters. Nevertheless, in terms of model construction, the current methods fail to address the noise heterogeneity effect and the spatial correlation effect in abundance coefficients, and in terms of model optimization, they fail to consider the purified means effect of the endmembers and the fuzzy class membership effect of the abundances coefficients, causing compromised model efficiency and difficulties in model optimization. The project therefore focuses on addressing the noise heterogeneity effect and spatial correlation effect for building enhanced linear spectral model, and focuses on building an efficient optimization scheme using fuzzy clustering approach, including, (1) the construction of heterogeneous noise model; (2) the modeling of spatial correlation effect; (3) the use of endmembers as purified means for avoiding complex prior distribution of endmembers; (4) the use of abundance coefficients as fuzzy class membership for building a fuzzy-clustering-based optimization algorithm. From a theoretical perspective, the proposed model provides new ways for modeling and optimizing the spectral unmixing issue. From an application perspective, the proposed model not only can be used for enhanced spectral unmxing and endmember extraction, but also can be used for other hyperspectral image processing tasks, such as denoising, feature extraction, clustering analysis, target detection and compression.

高光谱混合像元分解一直是研究热点和难点，关键在于光谱混合模型构建与参数估计，目的是获取高精度的参数估值并应用。目前方法在模型构建上没有考虑噪声的异方差性和丰度的大尺度空间相关性，在参数估计上没有考虑端元的净化均值特征和丰度的模糊标签效应，导致模型参数估计精度损失和优化计算困难。.本项目引入异方差和空间相关性以构造新的线性光谱混合模型，并研究基于模糊聚类优化该新模型的方法，研究包括：（1）噪声异方差模型的构建；（2）大尺度空间相关性模型的构建；（3）用净化均值替代端元复杂先验分布的优化算法；（4）将非负丰度系数作为模糊标签以形成模糊聚类算法；理论上，是对提高高光谱遥感数据建模精度并优化算法的贡献；应用上，该模型不仅可用于非监督混合像元分解、端元提取，并提高精度，而且可以用于高光谱去噪、特征提取、空间聚类、目标探测、图像压缩等其它高光谱图像处理，具有较广的实用价值。

高光谱混合像元分解是从混合像元中提取基本物质元素的光谱(端元)，并且求出这些物质所占的比例(丰度)，具有重要的理论和应用意义，一直是研究热点和难点。高光谱混合像元分解的关键在于光谱混合模型构建与参数估计,目的是获取高精度的参数估值并应用。目前方法在模型构建上没有考虑噪声的异方差性和丰度的大尺度空间相关性,在参数估计上没有考虑端元的净化均值特征和丰度的模糊标签效应,导致模型参数估计精度损失和优化计算困难。针对这些问题，本项目引入了异方差和空间相关性以构造新的线性光谱混合模型,并研究基于模糊聚类优化该新模型的方法。重点研究了以下内容:(1)噪声异方差模型的构建;(2)丰度空间相关性模型的构建;(3)用净化均值替代端元复杂先验分布的优化算法;(4)将非负丰度系数作为模糊标签以形成模糊聚类算法;(5)模型在不同高光谱数据处理任务中的应用。该研究提高了高光谱遥感数据建模及优化算法的精度和效率。提出的模型不仅可用于非监督混合像元分解、端元提取,并提高了精度,而且可以用于增强高光谱去噪、特征提取、空间聚类等其它高光谱图像处理任务的效率和精度,具有较广的实用价值。