Introduction¶
Prospectivity mapping is a widely used technique in mineral exploration. It employs a collection of machine learning tools to identify or exclude areas with potential for mineralization. This task is performed by data scientists through a programmed workflow, involving iterative steps of analysis and adjustment.
Machine learning tools are statistical tools consisting of algorithms with internal parameters that establish rules. These rules are determined during a training session, which relies on the probabilistic distribution of the data. The output of the algorithm is a probability for each point, indicating its likelihood of being mineralized. This probability is calculated by the model, and its accuracy depends directly on the quality of the trained model and the decisions made earlier in the workflow.
Prospectivity mapping sits at the crossroads of various specialties. It demands knowledge of data science tools, including which tools to use, in what sequence, how to implement them, and an awareness of their potential pitfalls. An understanding of statistics is crucial, encompassing their requirements, biases, and strategies for mitigating these biases. Geology plays the major role, guiding the choice of properties to use, identifying what is being searched for, and determining what is geologically plausible.
The iterative adjustment process is thoroughly informed by the interplay of these specialties: data-science, statistics, and geology. Data scientists continuously review and adjust the workflow’s parameters, tailoring them based on the outcomes at each step and integrating insights from geology. The approach is refined repeatedly, ensuring that the statistical methods employed effectively counteract biases while leveraging the correct geological indicators.
The goal of the Targeting-Workflow application is to provide geologists, including those without programming skills, with a robust set of machine learning tools tailored to support geological decision-making. This workflow facilitates an iterative process that reduces the risk of statistical biases, offering various methodologies to address them effectively.
Technically speaking, targeting in mineral exploration is a binary classification problem, where the goal is to differentiate between two categories: mineralized and non-mineralized areas. This process involves several challenges:
The binary classification takes place within a spatially auto-correlated dataset, where neighboring data points share properties, potentially leading to overfitting with machine learning techniques.
The dataset is often heavily unbalanced, with known mineralized points being scarce compared to the majority of non-mineralized areas, which can introduce statistical bias if not properly addressed.
The existence of a correlation between mineralization and certain properties does not imply causation, a situation that becomes especially problematic when few known positive targets exist.
To avoid these pitfalls, users must carefully review which properties the models use for their predictions and select or deselect properties based on geological expertise.
The workflow is designed to navigate these statistical challenges and simplify the use of such tools for geologists, enabling them to quickly apply their expertise to the process. Consequently, the targeting workflow is applicable to any heavily imbalanced “binary classification” task within an auto-correlated dataset, not just for mineral targeting.
Throughout the documentation, an example based on the “Flin Flon deposit,” available with the installation of Geoscience ANALYST, will be used to illustrate the workflow. The prospectivity analysis will be conducted on the surficial sediments to search for gold mineralizations.