Online Cluster Analysis Supporting Real Time Anomaly Detection in Hyperspectral Imagery.

摘要

Ongoing work in anomaly detection in hyperspectral images has shown that cluster analysis performed in appropriate principal component subspaces can enhance the performance of detectors such as the Reed-Xiaoli detector and its derivatives. Numerous operational considerations motivate the development of an online or incremental clustering algorithm, which can perform clustering as pixels of the image are collected in real time rather than waiting until the full image is complete. Such an algorithm is developed by combining key elements of existing clustering algorithms from related domains. The parameters of the algorithm are tuned and performance of the algorithm is assessed using a set of actual hyperspectral images by exploiting key attributes of an appropriate principal component sub-space. A byproduct of the clustering algorithm is a rudimentary anomaly detector which demonstrates the feasibility of cluster based outlier detection in hyperspectral imagery.