GRSG 35th Conference 2024 Abstract

Title: Rare Earth Element quantification using Random Forest regression and Reflectance Spectroscopy

Author: Susanna Grita

Organisation: University of Rome “La Sapienza”

Interest in Rare Earth Elements (REE) has increased in recent years since they are considered critical raw materials for Europe’s economy. In reflectance spectroscopy, the spectral signatures of these elements are characterized by sharp and narrow features in the VNIR and SWIR ranges. While it is known that the depth of these features is correlated to the concentration of REEs, only a few studies have attempted to investigate this relationship quantitatively. Here we test a machine-learning method for quantifying REEs using reflectance spectra.

A Random Forest regression model was trained by about 100 spectra of natural and synthetic REE-bearing samples to model the relation between the spectra and REE concentrations quantitatively. Preliminary results indicate an R2 of 0.91 for Nd and an RMSE less than 10% of the total training data range. A comparison between the model’s feature importance and the REE-related spectra shows that the Random Forest regressor correctly picks the diagnostic absorption features for developing its decision trees, proving to be a sound method for REE quantification.

The model was also tested on image cubes obtained from natural samples in the lab. The predicted average concentrations of Nd in the image cubes were in line with those reported previously using analytical geochemistry. The promising results obtained in this work can serve as a starting point for future studies on REE quantification using hyperspectral remote sensing technology.