A methodology for detecting image anomalies in a target area for classifying objects therein, in which at least two images of the target area are obtained from a sensor representing different polarization components. The methodology can be used to classify and/or discriminate manmade objects from natural objects in a target area, for example. A data cube is constructed from the at least two images with the at least two images being aligned, such as on a pixel-wise basis. A processor computes the global covariance of the data cube and thereafter locates a test window over a portion of the data cube. The local covariance of the contents of the test window is computed and objects are classified within the test window when an image anomaly is detected in the test window. For example, an image anomaly may be determined when a matrix determinant ratio of the local covariance and the global covariance exceeds a probability ratio threshold. The window can then be moved, e.g., by one or more pixels to form a new test window in the target area, and the above steps repeated until all of the pixels in the data cube have been included in at least one test window.