Methods for classification of hyperspectral imagery

Remote imaging using air or spaceborne sensors has become one of the most advanced tools for collecting information about the surface cover of the Earth and other planets. Possible uses of such data includes mineral exploration, land use mapping, forest management, assessment of natural hazards, estimation of water resources and detection of environmental contamination such as oil spills. Methods and tools developed for imagery obtained from observation of the earth can also be used to search for past or present life on other planets.

By introduction of the imaging spectrometer the amount of available information in remote imagery has been increased 100-fold. While traditional remote sensing platforms such as the Landsat series samples the visible and near-infrared spectrum of 0.4 - 2.5 um in several broad bands, hyperspectral sensors sample the spectrum contigously with very narrow bands.

My research revolves around the problems and blessings of the high dimensionality of imaging spectrometer datasets. While adding informative features about a object intuitively seem like a good idea, in practice problems quickly arise due to data sparsity. (The number of samples needed for reasonable parameter estimates grows extremely quickly in several classification methods.)

I focus on optimizing image analysis performance by feature extraction or utilizing ancillary/prior knowledge.