AI-Driven Permafrost Mapping Using Airborne Electromagnetic Data in Yukon, Canada

Sammanfattning

This study developed a comprehensive workflow for the understanding of permafrost distribution and related subsurface structures using airborne electromagnetic (AEM) data in Yukon, Canada. The framework combines geophysical inversion, sensitivity analysis, geological joint interpretation, and machine learning techniques to improve the characterization of permafrost in a remote region with limited data availability. A one-dimensional inversion model from the SimPEG library was applied to a large AEM dataset to generate resistivity values. Validation using two electrical resistivity tomography (ERT) profiles and borehole data indicated good agreement on subsurface characterization between AEM, ERT, and borehole, supporting the reliability of the inversion results. Sensitivity analysis and depth of investigation (DOI) assessment were included to evaluate model performance. An opacity control method driven by sensitivity values was developed on resistivity profiles to enhance visualization. Through the interpretation of resistivity profiles with several geological datasets, including surficial materials, lithology, vegetation, and topographic information, the distribution of permafrost, taliks, and other subsurface features were identified along the selected flight lines. A convolutional neural network (CNN) autoencoder was further applied to extract latent features from resistivity profiles, and clustering methods were used to identify potential patterns. The results demonstrate the potential of machine learning approaches to support semi-automatic interpretation of large AEM datasets. Overall, the proposed workflow integrates AEM inversion, sensitivity analysis, validation, geological analysis, and machine learning into a comprehensive framework for permafrost understanding. This methodology improves both the reliability and efficiency of subsurface interpretation and provides a baseline for future regional-scale permafrost learning and automatic analysis of geophysical data.