Estimation of Surface Reflectance and Mineral Composition by Combining In Situ and Remote Spectroscopic Measurements

Abstract

Remote mapping of surface mineral composition often involves Visible to short wavelength infrared (VSWIR) and Thermal Infrared (TIR) imaging spectroscopy. On the one hand, orbital multiband spectrometers - such as ASTER and LANDSAT - provide global-scale coverage and valuable contextual information; nonetheless, their low spectral resolution is often insufficient for mineral identification. On the other hand, there exist airborne instruments - such as AVIRIS-NG and PRISM - that count with a resolution that may be comparable to laboratory or ground spectrometers; however, this kind of observations are usually not available at a global-scale. In this work, we introduce a novel spatio-spectral machine learning model that estimates surface reflectance. It uses a low-resolution orbital prior that is updated every time a high-resolution in situ (ground) spectrum is collected. This model is pre-trained using both airborne and laboratory spectra. Furthermore, it is capable of building mineral composition maps by comparing the estimated reflectance to endmembers in a spectral library. We present a case study that consists in mineralogical investigations at Cuprite, Nevada by combining ASTER orbital measurements and ASD ground spectra. We utilize AVIRIS-NG surface reflectance, as well as the USGS spectral library, for pre-training and validation. Our results show that combining orbital and in situ spectra is an effective way to estimate surface reflectance and map mineral composition whenever high-quality airborne observations are unavailable.