Enhancing Deep Learning-based Strontium Isotopic Landscape Estimation Using Geostatistical Method: A Case Study in South Korea

This study introduces a novel methodology to enhance the accuracy and efficiency of ⁸⁷Sr/⁸⁶Sr isoscape mapping by integrating deep learning (DL) techniques with geostatistical methods. Utilizing kriging-based data augmentation, the proposed framework addresses data scarcity by generating high-quality synthetic training data while incorporating spatial geological factors and geochemical elements as input variables to improve model performance. The study developed an isotopic basemap for South Korea using 409 soil samples and a feedforward deep neural network (FDNN) model. The FDNN model demonstrated superior accuracy (91.67%) compared to traditional kriging (76.8%) and convolutional neural network (CNN)-based models (86.14%). The robustness of the FDNN model was significantly enhanced by kriging-based data augmentation, which not only captured geological anisotropies but also incorporated uncertainty analysis to improve reliability. The resulting ⁸⁷Sr/⁸⁶Sr isoscape map revealed distinct isotopic distributions across South Korea, with higher ratios associated with metamorphic and granitic rocks, reflecting geological history and topographical influences. Notably, the predicted isotopic distributions closely aligned with the boundaries of tectonic provinces, underscoring the geospatial accuracy of the developed model. Validation using bone samples additionally confirmed the efficacy of the proposed method in accurately estimating isotopic levels. These findings highlight the potential of combining geostatistical and DL approaches to overcome traditional challenges in isotopic mapping, offering scalable solutions for applications in environmental monitoring, archaeology, and provenance studies.