GRSG 35th Conference 2024 Abstract

Title: Hyperspectral Imaging and Predictive Model Evaluation for Total Organic Carbon Content Estimation: Bituminous Shales: A Case Study on Rocks of the Horn River Basin, Canada

Author: Carlos Souza Filho

Organisation: University of Campinas

The use of hyperspectral imaging in predictive models for the parameterization of geochemical and geological variables can provide various types of mineralogical and compositional information from rock samples. When successfully applied, it can reduce costs in rock core analysis processes. This study examines a 180-meter drilling core that intersects the bituminous shales of the Horn River Group, Canada. The aim is to assess the possibility of predicting Total Organic Carbon (TOC) using hyperspectral imaging data in the shortwave infrared region.

Machine learning algorithms such as Support Vector Machines and Convolutional Neural Networks provided three robust predictive models for TOC estimation. The best-performing model presented a mean squared error of 0.72% between predicted and actual TOC values. A sensitivity analysis of the model indicates the influence of hydrocarbon spectral signatures, with at least four spectral regions showing characteristic absorption features. A predictive image of the TOC content of the studied core was generated, revealing two-dimensional TOC variability consistent with lithostratigraphic descriptions. The results indicate significant potential for hyperspectral imaging data in geochemical predictions and in assisting with sampling strategy development in bituminous shales.