high performance computing on graphics processing units: hgpu.org

Spectral unmixing is one of the most popular techniques to analyze remotely sensed hyperspectral images. It generally comprises three stages: 1) reduction of the dimensionality of the original image to a proper subspace; 2) automatic identification of pure spectral signatures (called endmembers); and 3) estimation of the fractional abundance of each endmember in each pixel of the scene. The spectral unmixing process allows sub-pixel analysis of hyperspectral images, but can be computationally expensive due to the high dimensionality of the data. In this paper, we develop the first real-time implementation of a full spectral unmixing chain in commodity graphics processing units (GPUs). These hardware accelerators offer a source of computational power that is very appealing in hyperspectral remote sensing applications, mainly due to their low cost and adaptivity to on-board processing scenarios. The implementation has been developed using the compute device unified architecture (CUDA) and tested on an NVidia TM GTX 580 GPU, achieving real-time unmixing performance in two different case studies: 1) characterization of thermal hot spots in hyperspectral images collected by NASA’s Airborne Visible Infra-Red Imaging Spectrometer (AVIRIS) during the terrorist attack to the World Trade Center complex in New York City; and 2) sub-pixel mapping of minerals in AVIRIS hyperspectral data collected over the Cuprite mining district in Nevada.