基于信噪比估计的波段选择与高光谱异常检测

Band Selection based on Signal-to-noise Ratio Estimation and Hyperspectral Anomaly Detection

有效的波段选择方法可以极大地提高高光谱图像处理速度的同时改善处理效果。为了自动判断低信噪比波段,提出了一种基于小波变换的图像信噪比估计(SNR estimation,SNRE)方法,利用小波变换后对角方向上的高频成分估计噪声方差并计算信噪比。将该方法分别结合基于方差和相关系数(V_COR)的最优索引指数、最大信息量(MI)、高阶矩(偏度或峰度)结合信息散度(K3_KL)等3种基于信息量的波段选择方法后选择波段。将这些改进后的波段选择方法应用于高光谱异常检测。实验结果表明SNRE预选波段结合MI和K3_KL选择波段用于异常检测能进一步提高检测精度。

Effective band selection algorithms can greatly improve the hyperspectral image processing speed and effect simultaneously.In order to automatically determine low signal-to-noise bands,a new image signal-to-noise ratio estimation(SNRE)method is proposed based on wavelet transform.Performing wavelet transform on each band image which is assumed to be only corrupted by additive Gaussian noise,and the mid-value of high frequency component of the wavelet transform is used to estimate the noise variance,then further to calculate the SNR.This method is then integrated with three band selection methods based on information such as optimal index factor defined by variance and correlation coefficient(V_COR),maximal information(MI)and high order moments(kurtosis or skewness)combined with the divergence(K3_KL)to select bands respectively.These improved methods are evaluated by experiments of hyperspectral anomaly detection.Experimental results demonstrate that SNRE combined with MI or K3_KL which can further improve the results of anomaly detection.