基于条件随机场的光谱相似性匹配高光谱遥感影像聚类方法

A spectral similarity matching classifier based on conditional random field for hyperspectral remote sensing imagery

高空间分辨率的高光谱遥感数据不仅能够获取地物近似连续的光谱曲线,还具有丰富的空间信息.传统的基于单像元的光谱匹配方法无法将这两种特征很好地结合.针对该问题,提出将条件随机场(CRF)模型引入光谱匹配方法.CRF模型通过构造像元邻域描述空间信息,解决了基于单像元光谱匹配方法仅考虑光谱信息的不足,实现了聚类过程中光谱和空间信息的融合;然而,传统CRF模型基于欧氏距离和马氏距离等相似性测度,无法适应于高光谱遥感影像的数据特征,因此利用光谱相似性测度改进传统CRF模型的相似性测度准则.实验证明,所提出方法能够有效解决传统光谱匹配方法结果的噪声问题,较好地保留了地物的形状特征,分类精度得到提高.

Hyperspectral remote sensing imagery with high spatial resolution can not only obtain the ap- proximately continuous spectral curve, but also contain abundant spatial information. Traditional spectral matching algorithms based on single pixel are not able to combine these two features. In order to solve the problem, this paper introduces conditional random field（CRF） model into spectral matching algorithms. CRF model presents spatial information by constructing neighbor system of the pixel to improve single-pix- el spectral matching algorithms which only concern spectral information, thus fulfill the fusion of spectral and spatial information in clustering. However, traditional CRF model is based on similarity measures such as Euclidean distance and Mahalanobis distance, etc. It can＇t be directly applied to hyperspectral re- mote sensing imagery. For this reason, this paper improves traditional CRF model by using spectral simi- larity measures. The experiment shows that the proposed algorithm can effectively solves the noise prob- lem of traditional spectral matching algorithms, and maintain the shape features of the objects. The classi- fication accuracy is improved.