Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2009 · 30 citations · 5 references
ComputingComputer ScienceComputer VisionMachine LearningRemote SensingAnomaly DetectionFeature DetectionImage ClassificationObject DetectionOutlier DetectionRemote Sensing Image AnalysisAnomaly AnalysisDetectorsFalse Alarm Rate AnalysisGraph-based Anomaly DetectionHyperspectral AnalysisHyperspectral Imagery AnalysisImage Anomaly DetectionTopological Data Model
Anomaly detection algorithms applied to hyperspectral imagery are able to reliably identify man-made objects from a natural environment based on statistical/geometric likelyhood. The process is more robust than target identification, which requires precise prior knowledge of the object of interest, but has an inherently higher false alarm rate. Standard anomaly detection algorithms measure deviation of pixel spectra from a parametric model (either statistical or linear mixing) estimating the image background. The topological anomaly detector (TAD) creates a fully non-parametric, graph theory-based, topological model of the image background and measures deviation from this background using codensity. In this paper we present a large-scale comparative test of TAD against 80+ targets in four full HYDICE images using the entire canonical target set for generation of ROC curves. TAD will be compared against several statistics-based detectors including local RX and subspace RX. Even a perfect anomaly detection algorithm would have a high practical false alarm rate in most scenes simply because the user/analyst is not interested in every anomalous object. To assist the analyst in identifying and sorting objects of interest, we investigate coloring of the anomalies with principle components projections using statistics computed from the anomalies. This gives a very useful colorization of anomalies in which objects of similar material tend to have the same color, enabling an analyst to quickly sort and identify anomalies of highest interest.
5