UNSUPERVISED MACHINE LEARNING AND REMOTE SENSING INTEGRATION FOR GEOLOGICAL MAPPING AND MINERAL DETECTION

This study introduces an integrated approach for analyzing multispectral satellite imagery to generate surface geologic maps and identify specific minerals across diverse types of terrain. The main objective was to develop an unsupervised machine learning algorithm that reduces the dimensions of the satellite data and maintains accuracy during the classification process. The integration of Principal Component Analysis (PCA) and K-Means clustering allowed the multispectral data and Digital Elevation Models (DEMs) to reveal geological patterns and groupings.

PCA was first applied to transform the multispectral and elevation datasets into their most significant components, maintaining spatial integrity, while lowering computational demands. Second, K-Means clustering was used to identify natural groupings of classes within the reduced dataset. These classes were intended to align with lithologic units, mineral-rich formations, and other geologic features. The workflow successfully captured subtle spatial variations that might be missed with conventional methods, especially when different principal components were used to enhance unique features.

To validate the algorithm, multiple areas with diverse geologic settings were used so that areas with different lithologic compositions and exposed surfaces would challenge the algorithm. Initial development began with preprocessing data from Wyoming, followed by assessments in New Mexico, Nevada, and Arkansas. This allowed for evaluating the performance of the algorithm and refining the classification process, while also making it possible for the identification of parent rocks containing key minerals and combining classes into formations. Accuracy had strong comparisons of 79% to 86% when evaluating the algorithm’s generated maps with existing geologic maps.

The methodology of combining PCA and K-Means clustering to the combination of satellite imagery and DEM’s is a flexible tool for remote sensing-based geologic analysis. Its ability to visualize classes and adapt across different terrains makes it valuable for applications in resource exploration, environmental monitoring, and geospatial studies. This work highlights the potential of integrating machine learning and remote sensing.