Unsupervised Ensemble-Kernel Principal Component Analysis for Hyperspectral Anomaly Detection

Nicholas Merrill, Colin C. Olson; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020, pp. 112-113

Unsupervised anomaly detection--which aims to identify outliers in data sets without the use of labeled training data--is critically important across a variety of domains including medicine, security, defense, finance, and imaging. In particular, detection of anomalous pixels within hyperspectral images is used for purposes ranging from the detection of military targets to the location of invasive plant species. Kernel methods have frequently been employed for this unsupervised learning task but are limited by their sensitivity to parameter choices and the absence of a validation step. Here, we use reconstruction error in the kernel Principal Component Analysis (kPCA) feature space as a metric for anomaly detection and propose, via batch gradient descent minimization of a novel loss function, to automate the selection of the Gaussian RBF kernel parameter, sigma. In addition, we leverage an ensemble of learned models to reduce computational cost and improve detection performance. We describe how to select the model ensemble and show that our method yields better detection accuracy relative to competing algorithms on a pair of data sets.