A New Tool for Evaluating Spectral Unmixing Applications for Remotely Sensed Hyperspectral Image Analysis

Hyperspectral imaging is a new technique in remote sensing that collects hundreds of images, at different wavelength values, for the same area in the surface of the Earth. For instance, the Airborne Visible Infra-Red Imaging Spectrometer (AVIRIS) instrument operated by NASAs Jet Propulsion Laboratory collects 224 spectral channels in the wavelength range from 40 to 250 nanometers using narrow spectral bands. The new generation of satellite hyperspectral instruments improves this spectral resolution even more, providing very detailed spectral information about ground cover materials. However, the spatial resolution of hyperspectral imaging instruments is still in the range of several meters per pixel. As a result, the pixels collected by an imaging spectrometer are likely mixed in nature. Spectral unmixing is a very important tool in remotely sensed hyperspectral data exploitation which aims at estimating the abundance of pure spectral components (called endmembers) in each mixed pixel. During the past years, many algorithms and models have been developed for endmember extraction and abundance estimation in remotely sensed hyperspectral images, thus making spectral unmixing a hot topic in the hyperspectral imaging literature. However, there is no clearly standardized data set for benchmarking the accuracy of spectral unmixing techniques. In this paper we present a novel tool for hyperspectral unmixing that includes most of these techniques. Also the tool includes a database of sinthetic images generated using random fractal patterns and a real dataset obtained by de AVIRIS sensor of the NASA Jet Propulsion Laboratory. The tool is allowed to perform all the steps of the unmixing chain (Estimating the number of endmembers, feature reduction, endmember extraction and linear spectral unmixing) and also allows the result analysis using several metrics such as the spectral angle distance (SAD) and the root mean reconstruction error (RMSE). The developed open-source tool and quantitative comparison of algorithms is expected to be of great interest to both algorithm developers and end-users of spectral unmixing applications.