Texture, shape and context for automatic target detection and classification in spectral imagery

This dissertation addresses the detection and classification of man-made objects, or targets, in spectral imagery. The exploitation of spectral texture, object shape, spatial context, and temporal context for detection and classification are examined.; Several data collection and demonstration programs are described. The Dark HORSE 1 (DH1) hyperspectral detection experiment is detailed. DH1 represents the first demonstration of real-time autonomous target detection and cuing from hyperspectral imagery. The Daedalus and SEBASS sensors, whose data are used throughout the dissertation, are discussed.; The use of multivariate Gauss-Markov random fields (MGMRFs) to model spectral texture is studied. New parameter estimation, spectral texture segmentation and anomaly detection algorithms are developed based on an unrestricted first-order isotropic MGMRF. The first demonstration of this model for spectral texture segmentation and anomaly detection in spectral imagery is described. The new algorithms yield segmentation and anomaly detection results superior to the previously available Gaussian spectral clustering (GSC). The influence of parameter estimation method and the number of segments is quantified. Stochastic boundary unmixing, a novel detection approach utilizing scene segmentation, spatial context, and linear spectral mixing is developed.; A novel automatic target classifier that exploits shape and spectral characteristics is demonstrated. The classifier uses GSC and competitive region growing to extract spatial and spectral object features that are classified by an artificial neural network (ANN). Discrimination between multiple object classes including excellent discrimination between natural and man-made objects is demonstrated.; The first reported demonstration of object level change detection (OLCD) in spectral imagery is discussed. This method uses anomaly detection, segmentation and competitive region growth to find anomalous objects in multiple images of a scene. A site model consisting of a database of image segmentation maps and detected object characteristics is constructed and updated with each new image. Changes in the scene, such as the arrival or departure of objects, are detected as each new image is compared to the site model. The elimination of false alarms based on their persistence across multiple images is demonstrated. It is also shown that the performance of the OLCD system improves as additional images are processed.