A Simple Absolute Prospectivity Model Using R and Feature Vectors: Introduction and Project Setup

This article is the first of a four-part series of articles that will be released on a weekly basis. Links to the full series are available below:

• Introduction and Project Setup
• Endowment Analysis
• Importing Mappable Geological Data and Analysing the Relationships to Mineralisation
• Mapping Absolute Prospectivity Using Feature Vectors

Introduction

As Project Manager of Project Murchison at SRK Exploration (https://bit.ly/3iSVPEx) I spend a lot of time thinking about how we can quantify prospectivity. Project Murchison has a specific aim to calculate prospectivity in an absolute way ("Absolute Prospectivity") rather than relative prospectivity that is used in the industry at present. We are doing this in Project Murchison by using the starting point of a Mineral Systems framework and statistically modelling the relationship of each component to mineralisation and creating a joint probability model as an update to Bayesian priors that are based on geological expertise. This article aims to introduce a simpler technique to arrive at Absolute Prospectivity than the main research path of Project Murchison, but which could still be useful for strategically managing exploration. This analysis was completed by myself outside of Project Murchison but utilising the same data. It is therefore hopefully illustrative of some of the problems that Project Murchison is trying to tackle.

The article runs through an example that focuses on gold mineralisation in the Eastern Goldfields of West Australia and utilises the freely available Geoscience of West Austalia ("GSWA") mineral systems datasets. These are the same datasets utilised in the Y4 Mineral Systems Project. For information on deposit location, size and grade a West Australia gold deposit database purchased from MinEx Consulting and compiled by Richard Schodde was used. However, GSWA also produces the 'Minedex' deposit database which is freely available and could be used for a similar purpose. It should be noted that West Australia was picked due to the richness of the data and maturity of exploration, but similar models could still be constructed for data poor areas with some preparation, though with resulting estimates having a higher degree of uncertainty. As exploration is highly reliant on uncertainty to derive project value (see my previous article here), such methods can be particularly useful in data poor areas where quantifying uncertainty is most critical. In virgin terranes assumptions will need to be made regarding potential endowment, these assumptions should draw on comparable terranes which have some exploration history.

The model utilises a feature vector method whereby the search area is broken down into 1km2 cells, and values from each dataset recorded for each cell. The data from each dataset must be turned into categorical data, this will be illustrated later in the article. All the data in each cell can then be aggregated into a feature vector for further analysis. In this case it resulted in an area with 178,362 cells each with a feature vector of length 19.

After constructing the feature vectors, the frequency of deposits of different types in vectors of different types can be determined and linked to the probability of finding a future deposit by assuming that the mineral deposits found in the future will have broadly the same relationships to geological features as those found in the past. This assumption may not hold perfectly when considering new exploration search space such as exploration at depth, and adjustments would need to be made or a different technique used.

To create the link of what may be found in the future we will need to estimate the number of deposits left to be discovered. This could be done using conventional endowment analysis such as the USGS method or Zipf’s law. However, the issue with such techniques is that they result in an estimate of all undiscovered deposits that may be found at some point in the future. Though generally useful, particularly for governmental institutions, this has a limited utility for an explorer as they work on a limited time frame and do not have forever to find a deposit. Instead, I shall use a concept developed within Project Murchison called 'Discoverable Endowment', that is an estimate of how many deposits of different kinds will be found in the next exploration time frame. This exploration time frame can be adjusted to be an explorer's or investor's own time frame, for example an explorer may want to match it to the time period of their licence.

Not only is this sort of endowment analysis more informative to an explorer, it also deals with the difficult issue of exploration efficiency when considering Absolute Prospectivity. As discoverable endowment is based on the assumption that exploration is a continuous process that generates discoveries, it also encompasses our inherent inefficiency such as the fact that explorers may miss a deposit that gets found by a later explorer, and that some deposits are no doubt missed altogether & can perhaps be considered 'Undiscoverable' for a given exploration paradigm. It is possible to observe in exploration history the opening of new paradigms where deposits that where undiscoverable in the old paradigm become discoverable in the new and a new population of deposits become available for discovery. Traditional endowment cannot accommodate the issues of limited timeframe, exploration inefficiency and paradigm shifts in exploration search space.

Absolute Prospectivity is achieved by the synthesis of the probabilistic prospectivity models derived from the feature vectors and the Discoverable Endowment analysis. In this article this is achieved by simulations based on the feature vector model, followed by a sampling determined by parameters obtained from the Discoverable Endowment analysis. The final product is a method which can take a sub-area which may be for example an individual exploration licence and estimate a 'Absolute Prospectivity Value', as well as a probability density of the range of potential outcomes in terms of the magnitude of discovery/non-discovery and their value.

Project Setup

I conducted the analysis of this project in R Studio with only minor use of GIS software. Some datasets may require some manipulation before use, and this may be easier within GIS software that R. It is possible therefore to conduct similar analysis to this one solely with freely available software and packages (R Studio is free). An intermediate understanding of the R processing language is necessary, Python may be an alternative language for a similar outcome. Having a reasonably high-powered computer is helpful, particularly if you want to speed up processing times through parallel computing, I used a PC with a 2.6GHz i5 processor (4 core) and 8GB RAM. The project utilises large amounts of data and even larger files are generated during analysis, so plenty of storage and RAM are necessary. Significant processing time, occasionally running into days, is required for some of the computation. This is likely reducible using more efficient programming (I'm an Exploration Geologist rather than a programmer after all!), and/or additional computing power.

The data I used was either in .shp or .csv format and so could be imported into R directly using some of the standard geospatial packages. I used a number of R packages aside from the ubiquitous tidyverse and dplyr packages. At various points I have used each of the major spatial R packages including sf, sp, rgdal, raster & rgeos. For statistical model fitting I utilised the fitdistrplus package. No doubt with additional programming knowledge the number of packages utilised could be refined.

The first concrete step in the analysis is to define your projection, study area and grid size. I utilised the WGS 84 / UTM 51S projection to match my data and imported a predefined study area from a .shp file. It is important that the study area covers a reasonably coherent mineralised terrain as we are trying to assess the behaviour of gold deposits as a population, both in terms of their potential endowment and their relationship to the data. Note that ‘study area’ is differentiated from ‘area of interest’ which is the smaller subset of the mineralised terrane that is of particular interest to the user. It is also important to restrict the analysis to a single deposit class, in this case mesothermal gold. An analysis for multiple deposit classes will require each to be done individually and then combined afterwards.

I selected a 178,362km2 area that covers the entire Eastern Goldfields portion of the Yilgarn craton and is in fact the same area utilised for the Y4 project allowing for direct comparison. A 1x1km grid size was used as I consider this to be an appropriate scale at which we may plan and execute early to mid-stage exploration, and it also gives a good balance of precision vs speed. The grid size should relate to the potential size of a deposit, such that a 50x50m grid would be inappropriate as a mineral deposit could not be contained within a single square. Unfortunately, there isn't much flexibility to change this important parameter at a later stage, so it pays to spend a bit of time when selecting grid size as to what will work best.

This concludes the introduction and project setup and in the next step we be able to start look at the discoverable endowment as described in the introduction. This will be done by looking at the discovery deposits in a time series and will be covered in next weeks article 'Part 1: Endowment Analysis'