The identification of minerals by using reflectance spectra is an important issue since this technique is the most used by remote sensing campaigns. The spectral information can be wide depending on spectral range and resolution of the data acquired. This implies that the analysis can be complex and time consuming, often reducing hyperspectral data to a multispectral approach to investigate only some spectral features. A machine that helps to search a large variability in the features can be used into a first screening of a dataset; the goal of this task is to train a ML algorithm with the laboratory data in part of the visible to near infrared range that could be applied to spectra data coming from the planetary missions. This is a multi-features ML problem applied to a limited reflectance spectra acquired in laboratory activities during the last years.
Machine Learning techniques will be applied to mineral identification via reflectance spectra using some laboratory dataset between mineral endmembers and their mixtures, having as a variable the grain (or particle) size. We will train the machine to recognize specific spectral properties and their variability taking into account the limits of retrieving this information. Some set of spectra acquired between bi-modal mixtures, produced at more than one size range will be utilized to determine the optimal algorithm. The 2 end-members for each grain size will be used to teach the machine to recognize them and test the capability to define their presence in the mixtures, separating the sets of spectra from the same size range. Moreover we will include spectra from other cases to test the machine capability to distinguish these spectra and exclude them.