A NEURAL NETWORK APPROACH FOR SINGLE SCATTERING ALBEDO RETRIEVAL FROM THEMIS THERMAL INFRARED DATA

The Mars Odyssey Thermal Emission Imaging System (THEMIS) has been mapping the Martian surface in the thermal infrared (~6.8 to ~14.9 µm) since February, 2002, providing a “big data” archive for thermophysical and mineralogic studies with excellent global coverage. The thermal infrared and VNIR wavelength regions are complementary data sets for studying mineralogy because the former gives direct information on the fundamental vibrational modes of classes of minerals such as silicates and sulfates. We have developed a neural network approach to retrieve pixel-dependent temperatures and surface single scattering albedo (SSA) spectra from THEMIS IR data, using a first principles approach to solve this underdetermined problem, i.e., two independent variables, SSA and temperature. For a given THEMIS scene, we use DISORT to model radiative transfer from the surface, including effects of atmospheric gases and aerosols. To model surface SSA spectra, the Hapke bidirectional reflectance function is used as the surface boundary condition, with the emissivity taken to be the complement of the integral over the hemisphere, by Kirchhoff’s law. A large number of laboratory spectra for candidate rocks and minerals are combined with randomly chosen surface temperatures to generate synthetic THEMIS spectral radiance scenes. The synthetic scenes are then used to train neural networks to retrieve input spectra and temperatures, using the DISORT output to model radiances on sensor. Finally, the trained neural network is used to solve for actual spectra and temperatures. The methodology was tested using a THEMIS scene collected on October 1, 2013 while Curiosity was conducting surface measurements, including surface temperatures with the REMS instrument. The two data sets agree within several Kelvins and SSA values are consistent with mineralogy inferred from rover measurements.