Date of Graduation

Spring 2026

Degree

Master of Science in Geography and Geology

Department

School of Earth, Environment & Sustainability

Committee Chair

Kevin Mickus

Abstract

The accelerating global demand for lithium (Li) requires exploration strategies that can integrate heterogeneous geoscientific datasets, especially in structurally complex and poorly exposed terrains. The Ewoyaa Li deposit occurs in southern Ghana within the Paleoproterozoic Birimian Supergroup, where dense tropical vegetation and weathering limit bedrock exposure. This study integrates remote sensing, airborne geophysics, geological, and multi-element geochemical datasets to investigate Li prospects and to develop a machine-learning (ML) framework for exploration targeting. Medium and high-resolution Landsat-8, WorldView-3, and drone imagery were analyzed using spectral composites, band ratios, and principal component analysis to examine spatial and spectral patterns associated with the spodumene-bearing pegmatites and alteration zones. However, most likely due to dense vegetation and weathering, the remote sensing data showed limited effectiveness for mapping spodumene-bearing pegmatites and were not used in the final ML integration. Airborne magnetic datasets, consisting of magnetic RTE, derivative filtering, upward continuation, Euler deconvolution, and radiometric K/Th ratios, were analyzed to delineate structural patterns controlling pegmatite emplacement. The multi-element geochemical concentration dataset comprised 49 analyzed elements, of which 14 consistently available elements across soil, auger, and reverse circulation/diamond drill samples obtained using portable X-ray fluorescence (pXRF) were used. The geological, geophysical, and geochemical predictors, including granitoid and mafic dyke shapefiles, magnetic RTE, tilt derivative, and two K/Th ratio raster files, and 14 geochemical elements, were integrated within a spatially structured ML workflow using tree-based ensemble models (Random Forest, XGBoost, LightGBM, and CatBoost) evaluated using spatial GroupKfold cross-validation. The results show that among the tested models, LightGBM achieved the highest performance with PR-AUC = 0.70, ROC-AUC = 0.77, and a best-threshold F1 score of 0.64. The model interpretation using SHAP values indicates that radiometric K/Th ratios (KTH_MSU), geochemical indicators such as P and Rb concentrations, and structural predictors, including dyke complexity and Gaussian proximity to dyke systems, were the most influential variables controlling Li prospectivity. Approximately 85% of predictions within the top 10% of ranked areas correspond to known mineralized pegmatite occurrences, affirming strong enrichment of Li mineralization within the study area.

Keywords

Li, pegmatite, remote sensing, airborne magnetic and radiometric, ML, Ghana

Subject Categories

Data Science | Geochemistry | Geology | Geophysics and Seismology | Natural Resources and Conservation

Copyright

© Saadatu Abdullah

Download

Available for download on Tuesday, May 01, 2029

Open Access

Share

COinS