Object-level change detection in spectral imagery

Multitemporal monitoring of sites using spectral imagery is addressed. A comprehensive architecture is presented for the detection of significant changes in scene composition described at the object level of spatial scale. An object-level scene description is obtained by applying a statistical spectral anomaly detector followed by a competitive region growth object extractor. The competitive region growth algorithm is derived as the solution to an approximate maximum likelihood (ML) image segmentation problem. Gaussian spectral clustering is used to model the scene background. A digital site model is constructed that contains image segmentation maps and extracted object features. Object-level change detection (OLCD) is accomplished by comparing objects extracted from a new image to objects recorded in the site model. A restricted implementation of the architecture is described and tested on long-wave infrared hyperspectral imagery. It is demonstrated that spectral OLCD can eliminate false alarms based on their multitemporal persistence. Incorporating multiple images in the site model is observed to improve OLCD performance.