Laboratory Testing Of Mineral Detection And Abundance Algorithms With Visible-Short Wave Infrared Imaging Spectrometer Data

Abstract:

Testing advanced algorithms for mineral detection and abundance in the laboratory using mixtures of known abundance and particle size is crucial validation and verification work for analysis of remotely acquired data. Promising new approaches have emerged have not been rigorously tested in laboratory settings. Specifically, laboratory testing is required to evaluate the fidelity of detection algorithms with imaging spectrometer data that exhibit signal to noise, structured noise, and other properties characteristics of imaging spectrometer data sets (e.g. CRISM, M3). We show initial results of mineral detection methods (Factor Analysis/Target Transformation, FA/TT) and abundance determination with a Hapke nonlinear mixture model.

The experiments consist of custom sample trays of 2 cm x 2 cm boxes loaded with binary mixtures of a target mineral mixed with the Exolith Mars Global Simulant (MGS-1). We use target mineral abundances of 1, 2.5, 5, 10, 20 and 50%. Target minerals include gypsum, kaolinite, montmorillonite, nontronite, selenite, serpentine and calcite. Sample trays for a given target mineral are measured with a Headwall Imaging Spectrometer. The data have thousands of pixels/sample box providing ample statistical sampling for radiative transfer and statistical model assessments.

A simplified Hapke model provides abundance estimates of selenite in binary mixtures with MGS to within 5% for mixtures with 5% or more of selenite using the entire sample box. On a per pixel basis the estimates are highly variable. The mixtures with <5% selenite are not distinguishable from the blanks (pure MGS-1). We will compare this performance with other target minerals.

FA/TT is a statistical technique to identify the presence of minerals in hyperspectral data based on their spectral signatures. Eigenvectors from 10s to 1000s of pixels extracted from a hyperspectral imaging data set are linearly fit to a target laboratory spectrum. If the transform spectrum matches the target spectrum to a defined metric of measure (e.g. RMSE) then there is a high likelihood the material represented by that target spectrum exists in the hyperspectral data set. We will show progress towards determining the target mineral abundance detection limits, effects of structured noise and the rate of false positive detections.