Identification of pegmatite bodies, at a province scale, using machine learning algorithms: preliminary results

Abstract

Several raw materials for green energy production, such as high purity quartz, lithium, rare earth elements, beryllium, tantalum, and caesium, can be sourced from a rock type known as pegmatite. The GREENPEG project (https://www.greenpeg.eu/), started in May 2020, is developing and testing new and advanced exploration technologies and algorithms to be integrated and upscaled into flexible, ready-to-use economically efficient and sustainable methods for finding buried pegmatites and their green technology raw materials. One of the tasks of this project aims to apply different image processing techniques to different satellite images (Landsat, ASTER, and Sentinel-2) in order to automatically identify pegmatite bodies. In this work, we will present the preliminary results, regarding the application of machine learning algorithms (ML), more specifically, random forests (RF) and support vector machines (SVM) to one of the study areas of the project in Tysfjord, northern Norway, to identify pegmatite bodies. To be able to determine the classes that would make up the study area, geological data of the region, such as lithological maps, aeromagnetic data, and high-resolution aerial photographs, were used to define the four classes (1. pegmatites, 2. water bodies, 3. vegetation, 4. granite). All training locations were randomly selected, with 25% of the samples split into testing, and the remaining 75% split for training. The SVM algorithm presented more promising results in relation to overfitting and final image classification than RF. Testing the algorithms with several variables of parameters was able to make the process more efficient.