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.