Ponente
Descripción
This talk presents a machine learning framework for estimating FeO concentrations in the lunar regolith, focusing on a two-stage methodology that combines unsupervised and supervised learning. The approach addresses key challenges in hyperspectral data analysis, including high dimensionality, noise, and nonlinear relationships between spectral reflectance and composition. In the first stage, an unsupervised feature extraction process is applied to hyperspectral data fromthe Moon Mineralogy Mapper (M3). A multiscale wavelet transform is used to reduce noise and preserve relevant spectral structures, followed by a deep autoencoder that learns a compact nonlinear latent representation of the spectra. This step enables efficient dimensionality reduction while retaining the spectral information necessary for mineralogical interpretation. In the second stage, these latent features are used as input for supervised regression models to predict FeO concentrations. Several algorithms are evaluated, with Random Forest selected as the final model due to its robustness and ability to capture complex nonlinear relationships. The model is calibrated using laboratory spectra and geochemical data from Apollo samples and validated with independent observations. The proposed framework enables the generation of high-resolution, spatially consistent FeO maps and demonstrates strong generalization across different geological contexts. Overall, the combination of unsupervised feature learning and supervised prediction provides an effective and scalable solution for lunar geochemical mapping, with direct implications for planetary exploration and in-situ resource utilization (ISRU).