Nonparametric Framework for Detecting Spectral Anomalies in Hyperspectral Images

Over the past few years, hyperspectral data exploitation aimed at detecting spectral anomalies within remotely sensed images has been of growing interest in many applications. In this letter, we are interested in an anomaly detection (AD) scheme for hyperspectral images in which spectral anomalies are defined with respect to a statistical model of the background probability density function (PDF). Among the multitude of PDF estimators discussed in statistics literature, Parzen windowing (PW) has always attracted much attention. However, its ability to estimate the PDF of global background in order to detect anomalies in hyperspectral images has not been investigated yet. Here, we propose the use of PW to provide reliable background PDF estimation. As is widely recognized, PW performance is primarily affected by the choice of the bandwidth matrix, which controls the degree of smoothing for the resulting PDF approximation. In this letter, the bandwidth selection problem is approached, resorting to an unsupervised method based on a Bayesian approach. Once the background PDF is approximated through PW, it is employed to detect anomalous objects within the scene by using the likelihood ratio test.