Traditional hyperspectral anomaly detectors background statistics are easily contaminated by anomalies and not robust that is difficult to detect anomalies of nonlinear mixed. In response to these problems, the Kernel feature projection theory is utilized, robustness analysis method is introduced in the construction of the anomaly detector background information, and a robust anomaly detection method is proposed. Using this method, c hyperspectral data from lowdimensional space can be mapped to high dimensional feature space without specific nonlinear mapping function, the background and anomaly targets can be expressed by a linear model, and a robust detector is constructed in the feature space. This method reveals highorder features between the objects on ground surface and can reflect complex spectral characteristics of target and surface features distribution. Experiments of real hyperspectral images and simulated data can prove: 1) the proposed anomaly detection method has a better receiver operating characteristic (ROC) curve and area under the curve (AUC) statistics and has a greater degree of separation of the target and background; 2) in kernel feature space, exclusion of anomalies contamination on statistics of the background improve the detection accuracy; 3) feature extraction can make better utilizing spectral diversity distinguish anomalies and background, which is an important step of anomaly detector.