AbstractAdaptive Learning for Thermal Hyperspectral Unmixing Thermal Infrared (TIR) spectroscopy imager extends the range of hyperspectral sensors into the thermal infrared domain. Passive thermal infrared spectrometers for the Mid-Wave-InfraRed (MWIR 3.0-5.0 μm) and Long-Wave-InfraRed (LWIR 8.0-14.0 μm) wavelengths are readily available, thereby allowing wider and simultaneous wavelength coverage. In contrast to the visible region of the electromagnetic spectrum, the infrared region is dominated by thermal self-emission which allows the sensor to operate equally well during both day and night and to perform target detection and material classification regardless of illumination conditions. In this wavelength region, the measured radiance is a function of the abundances of materials, the optical properties of the materials but also of their temperature. This means that for N spectral bands there are N unknowns for the emissivity spectrum plus one unknown for the temperature. Therefore, spectral unmixing which is a common processing procedure in the reflective domain becomes complicated and unmixing methods that have been designed for the reflective domain are not suitable for the thermal infrared domain due to the spatial variability of the temperature and the joint dependence of the emissivity and the temperature. This talk presents two advanced new unmixing methods that estimate the abundance and the subpixel temperature in a mixed pixel by linearizing the black body law and by applying adaptive learning to estimate the subpixel emissivities. |