The feature mining of Hyperspectral remote sensing images for target detection, recognition and classification are divided into multi-pixels,single pixel and sub-pixel levels, which are more natural for target size scale and sensor's spatial resolution and easier for analysis of feature acquisition techniques. Differing from the traditional frame of only feature and feature extraction, a unified frame for the feature mining of Hyperspectral remote sensing images is summed up as (1) Feature selection which maintains the sensor' s physical meaning; (2) Feature extraction which comprehensively utilizes all the sensed data; and (3) Feature mixing which takes into account of all mixing information over a pixel covering multi small size (sub-pixel) targets. The research advance of Hyperspectral image feature mining is overviewed in a new and unified frame and its prospect is also outlined with pointing out some hot topics.%从目标空间尺度和传感器空间分辨率的相对大小,把高光谱遥感图像目标检测、识别与分类中的特征挖掘问题划分为多像元、单像元和亚像元3个层次,因而更具自然特性也更适合特征挖掘和目标分类与识别技术的分析.把高光谱遥感图像特征挖掘方法归纳为以保留波段物理意义为主要目的的特征选择、以综合利用所有观测数据信息为主要特色的特征提取,和考虑亚像元多目标混合信息的特征混合3大类.重点且简明地从高光谱遥感数据光谱曲线与光谱特征、特征提取、特征选择以及特征混合分析几个方面综述高光谱遥感数据/图像的特征挖掘技术的研究进展并通过热点问题展望其未来的发展趋势.
展开▼