Machine learning-based thorium normalization in airborne gamma-ray spectrometry for enhanced critical mineral exploration

Airborne gamma-ray spectrometry (AGRS), also known as radiometric surveying, is one of the most cost-effective large-scale methods for geological mapping, critical mineral exploration, and radiometric anomaly detection through the measurement of potassium (K), uranium (eU), and thorium (eTh) in rocks and weathered materials. However, the identification of radioelement anomalies is often affected by factors such as soil moisture variability, precipitation, and radiation attenuation. While primary single channel maps (K, eU, and eTh) reflect bedrock composition, secondary maps derived through mathematical operations among channels can highlight geochemical processes and modify radiometric contrast. Thorium is an important radioactive element in AGRS due to its relative geochemical stability and its relevance to critical mineral and nuclear energy-related systems and has therefore been widely used to normalize potassium and uranium variations. This study proposes a target-informed, machine-learning-based methodology for normalizing thorium concentrations using a Random Forest regression model to derive a model-estimated thorium reference response. The resulting reference response is then used to calculate relative normalized deviations and enhance thorium radiometric contrast in airborne gamma-ray spectrometry data. In this study, thorium is treated as a relatively stable reference component, and thorium-derived feature transformations, including squared thorium values and radiometric ratios (eTh 2 /eU and eTh/(eU + eTh)), are deliberately combined as part of a target-informed nonlinear normalization procedure. The squared thorium term emphasizes elevated thorium responses, whereas the ratio-based variables represent complementary relationships between thorium and uranium. The model was trained and evaluated using high-resolution AGRS data, and the resulting normalized products were analyzed in a Geographic Information System (GIS) to assess their geological relevance. The resulting model-estimated thorium reference surface response was used to derive normalized thorium values, modifying the spatial contrast of the original eTh response and highlighting spatially coherent thorium variations. Results indicate that the proposed feature transformations provide complementary radiometric information that may support geological interpretation and the identification of areas potentially associated with thorium enrichment and REE-bearing materials, demonstrating the potential of machine learning as an exploratory tool for radiometric data analysis and geological exploration.