Automated Classification of Visible and Near-Infrared Spectra Using Self-Organizing Maps

Existing and planned space missions to planetary and satellite surfaces produce increasing volumes of spectral data. Understanding the scientific content in this large data volume is a daunting task. Various statistical approaches are available to assess such data sets. We apply an automated classification scheme based on Kohonen Self-Organizing maps (SOM) developed originally for the thermal infrared (TIR) and extended here to the visible and near-infrared (VNIR). Available data from spectral libraries are used to train and test the classification in the VNIR. The library spectra are labeled in a hierarchical scheme with class, sub-class, and mineral group names. After training, the test spectra are presented to the SOM output layer and assigned membership to the appropriate cluster. These assignments are then evaluated to assess the robustness, scientific meaning and accuracy of the derived SOM classes as they relate to the spectral labels. We investigate the influence of particle size on our results by training and classifying three particle size separates. We find the SOM results are robust based upon the number of clusters determined from ten independent training/testing efforts. We find the SOM results are most scientifically meaningful for the grossest differences between materials, although some individual groups retain high accuracy even when the overall accuracy of the SOM is low.