高光谱影像小目标谐波分析探测模型

A Harmonic Analysis Model of Small Target Detection of Hyperspectral Imagery

针对高光谱影像小目标探测问题,结合白化处理(WP)与约束能量最小化(CEM)算法,提出一种基于谐波分析(HA)的小目标探测新模型,即HA-WP-CEM模型。该模型首先对原始影像进行3次谐波分析,提取出最适合小目标探测的高光谱影像谐波分析余项、振幅和相位等7个能量谱特征成分。其次通过对谐波分析特征成分的白化处理,实现各特征成分的背景抑制,突出低概率的小目标信息。在采集白化数据的探测目标参考像元光谱之后,将参考光谱矢量和白化数据协方差矩阵输入CEM算子,得到小目标的探测结果。最后利用高分辨率影像进行定性评价,并选取3个重点区域,通过计算模型探测结果的信噪比、探测率、漏检率和虚警率进行定量评价。研究结果表明,HA-WP-CEM探测模型在运行简易性、可靠性、实用性和准确性等方面均明显优于直接采用CEM、WP-CEM等方法。同时该模型在数据变换、抑制背景、降噪等方面也达到了很好的效果。

Combining with whitening processing（WP） and constrained energy minimization（CEM） algorithm,a new model on detecting some small targets of hyperspectral imagery,named the HA-WP-CEM model,is put forward based on the harmonic analysis（HA） technology.Firstly,the harmonic residual,amplitude and phase terms of the hyperspectral imagery are extracted by the harmonic analysis for 3 times,the seven energy spectral feature components of the harmonic analysis are well suited for detecting the small targets.Secondly,the WP technology is used to obtain the whitening data of the seven components,suppress the background information of the energy spectral feature components and highlight the prominent information of low-probability small target.Thirdly,the spectral vectors of referential pixels on detecting targets in the whitening data are collected and the covariance matrix of the whitening data is calculated.The spectral vectors and the covariance matrix are used as the inputs of CEM,and then the small targets of the hyperspectral imagery are detected.Finally,a referential high-spatial resolution image is used to assess qualitatively on the model,the signal to noise ratio（SNR）,the correct detection ratio（CDR）,the missed detection ratio（MDR） and the false alarm ratio（FAR）.The detecting results show that the HA-WP-CEM model is obviously better in the aspects of processing simplicity,reliability,practical applicability and accuracy than CEM,WP-CEM,PCA-CEM and PCA-WP-CEM.The model also can get good effect on transforming data,suppressing background and reducing noise of hyperspectral imagery.