Tackling Spatial Multiple Features AI/ML Problems in Geology with Hexagons

As Artificial Intelligence (AI) and Machine Learning (ML) methods evolve at an explosive pace, there is an increased need to handle geological data challenges if we wish to (continue to) ride the AI/ML wave. Most geological data is at its core geospatial data in different shapes and formats. A few examples are polygon-based geological maps, geophysical and remote sensing raster grids and a plethora of sample analyses with coordinate data. Applications of geological data are as varied as the Earth is vast. However, these differing geospatial data formats in geology significantly limit the interoperability of datasets in an analytical ML context. Multivariable analyses of geological data often involve extensive spatial interpolation and projection headaches. Consequently, we must first solve geospatial data challenges to fully tackle inter- and intradisciplinary geoscience problems with ML and AI predictions on multivariable cross-disciplinary geological data. At our company, Scandinavian Highlands, we are building a platform and database structure to break down these geospatial format barriers using a hexagonal discrete global grid system called H3. The H3 grid represents all positions on Earth’s surface by hexagon (and 12 pentagon) cells at different levels of coarseness, ie. resolutions, down to 1 m² cell area. The resolutions are bound together by a systematic parent cell to children cells hierarchy. We process different types of geospatial geological data (raster and vector) to an H3 grid representation at the appropriate resolution for the given dataset. In doing this, we create a database structure where different geological data layers can be seamlessly merged into a single feature “stack” table for AI/ML purposes at either local, regional or global scales and across individual dataset resolutions. In this presentation, we demonstrate the hexagonal multiple feature stack concept in action, from simple grouped/filtered visualisation, regression and descriptive statistics to dimension reduction techniques (e.g. PCA and t-SNE), clustering and other supervised and unsupervised methods. Furthermore, all analysis results can be assessed spatially on the map, grounding them on the Earth’s surface and in real-life decision-making use cases.