基于最速上升算法的超光谱图像波段选择搜索算法

Band selection algorithm of hyperspectral image based on steepest-ascent search algorithm

超光谱遥感数据具有的波段数目多、波段宽度窄、数据量庞大等特点,给图像的进一步解译带来困难。结合超光谱图像波段选择的具体应用,根据波段之间的相关性将整个波段划分为几个子波段,采用最速上升的特征选择搜索算法在各子波段中快速提取最优波段。为了验证本算法的有效性,分别选取JM距离、BH距离以及类内类间离散度作为评价准则,针对一幅200波段的AVIR IS超光谱图像进行分类实验,并将该方法与传统的SFFS算法进行对比。实验结果表明所采用的算法用于特征选择具有搜索能力强、分类精度高的特点,完全可以替代传统的SFFS算法。

This paper proposed a band selection approach of hyperspectral image based on steepest-ascent search algorithm. The approach needed to divide the whole hyperspectral band into several subgroups in terms of the relativity between bands firstly, and then applied the steepest-ascent search strategy to quickly extracting optimal band in every subgroup in which the combinations of bands was indicated by binary vectors and the search was being along the steepest direction until the local extreme was acquired. In order to verify the ralidity of this algorithm, the approach was compared with the classical sequential forward floating selection suboptimal techniques, using hyperspectral remote sensing images as a data set. Experimental results prove the ralidity of this algorithms, which can be regarded as a valid alternative to classical SFFS method.