A Kurtosis-Based Test To Efficiently Detect Targets Placed In Close Proximity By Means Of Local Covariance-Based Hyperspectral Anomaly Detectors

This paper focuses on the detection of targets placed in close proximity by means of local covariance-based anomaly detectors. Specifically, RX algorithm is considered as a case-study in order to show how covariance corruption due to target signal contamination within local background pixels can be mitigated by means of robust sample covariance matrix estimators. Contrary to previous works, where the heavy computational complexity of robust covariance estimator has prevented its local application or required a too high computational demand, here robust covariance estimation is selectively applied only on those image pixels most susceptible to covariance corruption. This is achieved by performing a quick local test at each pixel based on the sample kurtosis. Real data are employed to give experimental evidence of the performance provided by the proposed AD strategy in terms of both detection and computational efficiency.