Abstract:

Anomaly detection of remotely sensed data is one problem of the application research fields paid more attention to. It has potential applications in many fields such as military object recognition and environmental protection. Without loss of generality, it can be assumed that background pixels of remotely sensed data distribute within a series of space-varying hyper-ellipsoids while anomalous pixels locate out of those hyper-cubes. Holding on this hypothesis, the authors first apply Weiszfeld method to estimate the centers and cross-band covariances of these hyper-ellipsoids in remotely sensed data, then compute the Mahalanobis distance of each pixel to the center of the corresponding hyper-ellipsoid and determine the threshold using Mahalanobis histogram, finally, recognize the anomalous pixels of which the Mahalanobis distance is over the threshold. The authors develop a visual C++ program for recognizing anomalous pixels from remotely sensed data on the basis of GDAL function library for input and output of remotely sensed data. The authors conduct an experimental application on the new methd using the TM image of Atlanda. The experimental results show that the method can properly recognize local anomalies in remotely sensed image data.

Key words: Weiszfeld method, Mahalanobis distance, covariance matrix, anomalous pixel extraction, remotely sensed images