基于多重分形参数的高光谱数据特征提取

针对单一分形维数不能表征高光谱数据光谱局部吸收特征的问题,提出了基于光谱概率测度的多重分形参数特征提取方法.基于光谱信息度量进行光谱概率测度的计算,基于配分函数法估计得到尺度函数;通过对尺度函数求导计算出Holder指数,并对尺度函数勒让德Legendre变换计算出多重分形谱;从多重分形谱和Holder指数之间的函数关系提取表征多重分形谱形态的4个多重分形谱参数作为光谱特征参数;并应用于基于最小距离准则的航空推扫式高光谱成像仪(PHI,Prush-broom Hyperspectral Imager)图像监督分类.结果证明:利用基于光谱概率测度的多重分形参数特征提取方法提取的光谱特征参数进行分类得到的总体分类正确率达94.789%,分类精度明显高于利用信息量维数和多重分形谱特征提取方法进行分类的结果,证明了基于光谱概率测度的多重分形参数特征提取方法提取的多重分形参数的有效性和可靠性.

Feature extraction method based on multifractal parameters for hyperspectral imagery

Multi-fractal parameter extraction method based on spectral probability measurement was proposed to resolve the problem that the local absorption characteristics of hyperspectral data can not be described by the single fractal dimension.The method of spectral information measurement was used to calculate the spectral probability.The scaling function was estimated with the partition function.The differential coefficient of scaling function was calculated to obtain Holder exponent,and the multi-fractal spectrum was computed with Legendre transformation of scaling function.Four multi-fractal parameters can be extracted from multi-fractal spectrum and Holder exponent.The minimum Euclidean distance rule with the characteristic extraction based on multi-fractal parameters was applied to hyperspectral image supervised classification.The hyperspectral image was collected by airborne push-broom hyperspectral imager(PHI).The applied results show that the efficiency and reliability of the proposed method and its classification accuracy are about 94.789%,which is better than the classification accuracy of information fractal dimension and multi-fractal spectrum.