Emissivity retrievals from airborne infrared hyperspectral images coupled with visible to SWIR hyperspectral images

Emissivity extraction from hyperspectral infrared data is a challenge as it is a under-determined problem, and particularly with noisy data. A new method for emissivity retrieval dedicated to hyperspectral data that cover infrared as well as reflective spectral domains has been developed to overcome difficulties related to noise. The method is derived from the EARTH (surface Emissivity, temperature and Atmosphere Retrievals from Thermal infrared Hyperspectral image) algorithm [1], but takes advantage of the abundance of spectral information in the reflective domain, by including a classification step in the emissivity retrieval process. The attributes considered in this step are the first principal components of the reflective image (in radiance or reflectance units) that contain rich spectral information, and the brightness temperatures computed on the infrared image after spectral binning that give information on surface temperature. A k-means classification is carried out. Infrared spectral radiances are then averaged in each class and atmospheric compensation, based on neural networks followed by surface temperature / emissivity separation (TES), based on spectral smoothness is achieved. The benefits of the classification step are demonstrated on AHS data. Then the method is applied on simulated noisy data. The accuracy on emissivity is clearly improved when compared to results obtained without classification step.