dennis will be on hand at the poster session to discuss his latest work with eric grunsky on the application of machine learning to mineral deposit prediction using regional stream sediment data
Eric and Dennis will be presenting a summary of their latest project sponsored by Geoscience BC. The title of the poster is (and it's a mouthful!): Mineral Resource Prediction Using Advanced Data Analytics and Machine Learning of the QUEST-South Stream Sediment Geochemical Data, Southwestern British Columbia.
Geoscience BC conducted an infill stream sediment sampling program in the QUEST-South project area in 2009 and reanalyzed archived regional geochemical survey (RGS) samples using ICP-MS in 2010. Catchments were determined for these samples in 2011 and a preliminary interpretation of the geochemical data undertaken using the dominant rock type in the catchments to level the data for the effects of variable background. In this new Geoscience BC project we interpret the data from 8545 samples using the random forest procedure. Data for 35 elements were levelled for laboratory analytical effects and values below the lower limit of detection imputed prior to a centred log ratio transformation to moderate the effects of geochemical closure. Principal component analysis and t-distributed stochastic neighbour embedding (t-SNE) were used to reduce the number of variables required to enhance the geochemical signals associated with mineral deposits. Each sample was also attributed with the closest MINFILE occurrence, excluding anomalies and showings, within 2.5 km of the sample site. MINFILE occurrences were grouped based on similarities in BCGS mineral deposit models and geochemical signatures for training purposes. A training data set of 474 samples, including 100 samples not attributed with a MINFILE occurrence, was used to generate random forest posterior probabilities for the remaining 8071 samples based on the most significant principal components. The posterior probabilities for various grouped mineral deposit models have been used to generate kriged images to test for geospatial coherence in the predictions. Catchment polygons thematically coded using the posterior probabilities provide maps of exploration potential.