Mineral potential mapping of the REE-Line of Bergslagen, Central Sweden using random forest classifier modelling: application and testing of the EIS toolkit for prospectivity mapping.

The ongoing green transition and growing need for European raw material independence requires the development of innovative exploration methods, including mineral prospectivity mapping using machine learning (ML) based methods. The Exploration Information Systems project (Horizon Europe grant No. 101057357) has developed an open-source GIS plugin to enable user-friendly generation of mineral potential maps using ML using a mineral systems modelidentifying key evidentiary factors for mineralisation such as source, trap, modification and transport (McCuaig et al, 2010). Using a mineral systems model developed through studies of the REE-Line in Sweden, a prospectivity map has been generated using the random forest modelling tool within the EIS toolkit to test the toolkit and to identify potential new prospective areas for REE mineralisation.

REE mineralisation in Bergslagen, Sweden occurs primarily within skarn type polymetallic deposits formed within supracrustal carbonate layers, intercalated in metasupracrustal rhyolitic units dating from between 1.92-1.88 Ga. Key factors found to be favourable for REE mineralisation used as evidentiary layers in the RF model were: proximity to magnetic anomalies (source), principal components of geochemical signatures from glacial till (modifying processes), distance to linear structures and kernel density of linear structures (transport), evidence of K-Mg-Ca-Fe-Na alteration in bedrock and presence of carbonate/skarn horizons (trap) (Andersson et al., 2024).

RF models learning models require training data: i.e. points containing a deposit (1), or no deposit (0) to evaluate the probability any given pixel is to have a deposit. Due to the numerous occurrences of polymetallic deposits within the extent of the that lack REE mineralisations, two types of negative training points were used: mineralised polymetallic deposit with no REE occurrence, and non-mineralised points.

The final RF model demonstrated and a true positive rate of 59.02%, a false positive rate of 4%, and a true negative rate 36.6%. The RF model gave an area under curve of 0.97, demonstrating a probable overfitting of the data. This may be due to the somewhat smaller number of training points than is typically ideal for RF modelling (Carranza and Laborte, 2015). Additionally, the tight clustering of many of the training points may point to the need for a wider spatial distribution of positive training points to improve the model. The extent of mineral claims made by prospecting companies were overlain on the final model as a secondary validation of the map where good correlation was shown between the most prospective areas and current exploration. RF mapping of the REE line thus shows good potential, though improvements to training data are needed.

McCuaig, T. C., Beresford, S., & Hronsky, J. (2010). Translating the mineral systems approach into an effective exploration targeting system. Ore Geology Reviews, 38(3), 128-138.

Andersson, S. S., Jonsson, E., & Sadeghi, M. (2024). A synthesis of the REE-Fe-polymetallic mineral system of the REE-line, Bergslagen, Sweden: New mineralogical and textural-paragenetic constraints. Ore Geology Reviews, 106275.

Carranza, E. J. M., & Laborte, A. G. (2015). Random forest predictive modeling of mineral prospectivity with small number of prospects and data with missing values in Abra (Philippines). Computers & Geosciences, 74, 60-70.