Utilizing deep learning-based feature engineering for effective geographical origin subdivision and classification of environmental soil samples in South Korea

The application of complementary geochemical analysis alongside deep learning techniques serves as a powerful tool in identifying geographical origin of environmental samples on a national scale. This research presents methodologies aimed at enhancing the accuracy of origin determination for soil samples across South Korea, leveraging geochemical and geological data. It addresses challenges associated with integrating Sr isotopes and multivariate geochemical variables and preserving geological interpretability by incorporating Autoencoder deep learning algorithms,which facilitate efficient feature engineering for comprehensive data analysis. Through the analysis of 412 soil samples collected nationwide, a geographic origin distribution and classification model was developed, establishing a novel framework for environmental sample analysis. The analysis identified six origins within South Korea, each distinguished by its geological tectonic units, bedrock age, and bedrock type. Extensively wide areas with granite bedrock nationwide were mostly classified into the same origin, irrespective of their geological tectonic configurations. The findings highlight the efficacy of integrating isotopic with geochemical data through advanced analytical techniques, significantly improving origin tracing accuracy and efficiency. Such advancements have significant implications for disciplines including agriculture, forensics, and archaeology, showcasing the potential of these methodological innovations.