Noise Simulation and Correction in Synthetic Airborne TIR Data for Mineral Quantification

Abstract

Rock-forming minerals (such as feldspar and quartz) can be identified and quantified from thermal infrared (TIR) laboratory spectroscopy using spectral models. This paper uses synthetic airborne TIR spectra to test whether the hyperspectral Spatially Enhanced Broadband Array Spectrograph System (SEBASS) would theoretically be able to detect quartz and feldspar minerals and quantitatively predict mineral modes in felsic igneous rocks. Data from a previous laboratory study were used to simulate TIR spectra with band locations and noise levels of the SEBASS sensor. The quantitative partial least squares regression (PLSR) models from that study were applied to newly created synthetic SEBASS data, and results were compared with the predictions from the previous study. Predicted compositions based on SEBASS band positions are nearly identical $(\rho = 0.995)$ to those based on laboratory resolution. Results are still reliable [prediction errors within 0.4% (absolute)] to the original laboratory PLSR predictions when adding up to 1% noise (about five times the SEBASS noise level) to the synthetic data. Prediction errors rapidly increase when noise levels beyond 1% are used. These results show that SEBASS' spectral resolution, spectral coverage, and signal-to-noise levels are sufficient to quantitatively predict quartz and feldspar amounts, and feldspar compositions with models based on PLSR. Spectral distortions, such as reduced spectral contrast, tilts, and vertical shifts, must be compensated for before these quantitative models are applied. A mean and standard deviation (MASD) normalization is proposed using a set of ground data for compensating systematic errors that are common to all image pixels.