Machine learning for fully quantitative mapping of geology and environment

Northern Ireland soil chemistry

Geology has its roots in field observations made by the naked eye, but as technology has evolved, so too have the methods that allow us to map and understand the ground beneath our feet.

In recent years, there has been a trend towards the collection of ever-more quantitative data, a movement that has so far been exemplified by the Tellus projects in Northern Ireland and south-west England. These projects consisted of series of surveys collecting both geochemical and geophysical data through combined ground and air surveys.

The beauty and utility of these multidisciplinary surveys is uncovered not in the independent analysis of individual variables, of which there are many, but in the treatment of the large, multivariate datasets as a single system: a fully quantitative representation of the geo-environment as a whole.

Thanks to the current global explosion in big data analytics, we have at our disposal more tools than ever before to make sense of such complex data systems. We can use these tools — machine learning algorithms — to explore our data and to condense them into meaningful, evidence-based solutions for specific end users.

South-west England geochemistry
3D legend for the map in Figure 2a

Arsenic enrichment in south-west England

For example, the use of regression tree ensembles has allowed us to map the chemical composition of Northern Ireland (Figure 1) and south-west England (Figures 2a and 2b) with greater detail and accuracy than ever before, and with transparent conveyance of uncertainty. Building on this, neural networks have been used to map elemental enrichments with even better accuracy (Figure 3), providing valuable maps to guide environmental studies and mineral exploration.

The use of machine learning has so far been focused on modelling and predicting geochemistry; after all, chemical composition is a direct quantification of what a rock or soil actually is, and therefore provides an ideal data format through which to port traditional mapping workflows into the digital domain.

As interest grows in improving evidence-based understanding of the behaviour of the rocks beneath our feet we will need to augment our data systems with measurements of a wider range of geo-properties. For example, in addition to data on chemical composition, we can incorporate data on mechanical properties, hydrological properties and so on in order to provide reliable decision support tools for a wider user base.

For now, machine learning is successfully allowing us to translate multisensory, airborne-survey data into accurate, high-resolution maps of the chemical composition of the rocks and soils that support the foundations of our society.


Contact Charlie Kirkwood for more information.