基于选择性分段行-列二维主成分分析的高光谱图像异常检测

Anomaly Detection Based on Selective Segmentation Row-Column Two-Dimensional Principal Component Analysis for Hyperspectral Images

高光谱图像具有越来越高的空间和光谱分辨率,其带来了数据量大、相关性强和冗余度高的问题,使得异常检测结果精度不高。为了选择更加有利于异常检测的图像,运用二维主成分分析(2DPCA)方法降维,并引入局部联合偏度-峰度指数进行图像选择,提出了一种基于选择性分段2DPCA的高光谱图像异常检测方法。首先利用相关系数对原始图像进行分段,然后通过旋转数据结构在每个波段子空间中实现行-列二维主成分降维;再选择合适大小的窗口,遍历每个降维结果的主成分,计算窗口内的局部联合偏度-峰度指数,并以此为指标选择用于异常检测的图像。实验结果表明,所提方法的接收机工作特性(ROC)曲线、曲线下面积(AUC)值和Bhattacharyya距离值均优于其他传统的方法,因此具有更好的检测性能。

Hyperspectral image have higher and higher spatial and spectral resolution,resulting in a large amount of data,strong correlation and high redundancy,which makes the low accuracy of anomaly detection result.In order to select the image which is more favorable for anomaly detection,we use the two-dimensional principal component analysis（2DPCA）method to reduce the dimension,and introduce the local joint skewness-kurtosis index to image selection.A method based on selective segmentation 2DPCA for hyperspectral image anomaly detection is proposed.Firstly,the original image is segmented by the correlation coefficient,and then the row-column two-dimensional principal component dimension reduction is realized in each band subspace by rotating the data structure.Then,we select an appropriate size window to traverse all the principal components of each dimension reduction result.Meanwhile,the local joint skewness-kurtosis index is calculated in this window,which is regard as an indicator to select the image for anomaly detection.The experimental result shows that the receiver operating characteristic（ROC）curve,the area under the curve（AUC）value and Bhattacharyya distance value of the proposed method are better than other traditional methods,so that it has a better detection performance.