Design and evaluate a data-driven model for mapping subsurface lithology using remote sensing and geophysical data

Thesis Abstract

Advancements in subsurface geological mapping are critical for mineral exploration, groundwater management, and environmental assessment, yet traditional methods relying on point-based sampling and borehole data are often limited by high costs, spatial sparsity, and environmental constraints. This study addresses these issues by proposing a data-driven model that integrates remote sensing and geophysical data to enhance the accuracy and resolution of subsurface lithology mapping. The primary aim is to design, implement, and evaluate a predictive framework capable of classifying subsurface lithologies over extensive areas with minimal invasive procedures. Specific objectives include analyzing the correlation between multispectral satellite imagery, aeromagnetic, and gravity data with subsurface lithological units, developing a machine learning paradigm for lithological prediction, and assessing the model's performance relative to existing geological maps. The research adopts a quantitative, exploratory, and validation-oriented design, utilizing a combination of supervised machine learning algorithms—such as Random Forest and Support Vector Machines—for predictive modeling. The study population comprises lithological data pooled from a regional geological survey covering approximately 10,000 square kilometers characterized by diverse rock types, including igneous, sedimentary, and metamorphic formations. A stratified random sampling technique was employed to select 300 sample locations where existing borehole logs, core samples, and laboratory analyses provided ground-truth lithological labels. Remote sensing data, including multispectral imagery from Sentinel-2 and Landsat 8, were preprocessed for atmospheric correction and spectral enhancement. Geophysical data consisted of high-resolution aeromagnetic and gravity surveys acquired at spatial resolutions of 10 meters and 50 meters, respectively. Data collection involved GIS-based extraction of spectral indices such as NDVI, NDWI, and the Normalized Difference Built-up Index (NDBI), alongside spectral reflectance features. Geophysical measurements were processed through filtering and automated anomaly detection to delineate subsurface features. The primary analytical techniques employed encompass correlation analysis, principal component analysis (PCA) for feature reduction, and machine learning classification models, with model validation conducted via k-fold cross-validation (k=10) and assessment metrics including overall accuracy, precision, recall, and the F1 score. The study also employs ROC curve analysis to compare classifier performance, and feature importance measures to interpret model predictions. Expected findings suggest that the integrated model will significantly outperform traditional mapping approaches, with anticipated classification accuracies exceeding 85%, especially in complex geological terrains. The model is expected to reveal distinct spectral-geophysical signatures associated with specific lithological units, providing detailed lithological maps at a regional scale. Such results will demonstrate the potential of remote sensing combined with geophysical data to serve as a cost-effective and reliable tool for subsurface lithology prediction. This research contributes to the body of knowledge by developing a robust, reproducible framework that combines multiple data sources within a machine learning context for geological mapping. It advances the understanding of spectral-geophysical correlations and their applicability to lithological discrimination, filling existing gaps in remote sensing-based subsurface modeling. Additionally, the study proposes a scalable methodology adaptable to different geological settings, thus supporting resource management and environmental planning. The main conclusion underscores the efficacy of an integrated, data-driven approach in improving the spatial understanding of subsurface lithology. The study recommends further research into deep learning techniques, the integration of additional geophysical methods such as electrical resistivity tomography (ERT), and real-time data processing for dynamic geological monitoring. Overall, the findings advocate for a paradigm shift from traditional, labor-intensive geological surveys toward automated, remote sensing-based predictive modeling in subsurface geology.