Evaluating the performance of support vector machines (SVMs) and random forest (RF) in Li-pegmatite mapping: preliminary results

Machine learning algorithms (MLAs) have gained great importance in remote sensing-based applications, and also in mineral prospectivity mapping. Studies show that MLAs can outperform classical classification techniques. So, MLAs can be useful in the exploration of strategical raw materials like lithium (Li), which is used in consumer electronics and in the green-power industry. The study area of this work is the Fregeneda-Almendra region (between Spain and Portugal), where Li occurs in pegmatites. However, their smaller exposition can be regarded as a problem to the application of remote sensing methods. To overcome this, Support Vector Machine (SVM) and Random Forest (RF) algorithms were applied to. This study aims at: (i) comparing the performance accuracy in lithological mapping achieved by SVM and by RF; (ii) evaluating the sensitivity of both classifiers to class imbalance and; (iii) compare the results achieved with previously obtained results. For these, the same Level 1-C Sentinel-2 images (October 2017) were used. SVM showed slightly better accuracy, but RF was able to correctly classify a larger number of mapped Li-bearing pegmatites. The performance of the models was not equal for all classes, having all underperformed in some classes. Also, RF was affected by class imbalanced, while SVM prove to be more insensitive. The potential of this kind of approach in Li-exploration was confirmed since both algorithms correctly identified the presence of Li-bearing pegmatites in the three open-pit mines where they outcrop as well in areas where Li-pegmatites were mapped. Also, some of the areas classified as Li-bearing pegmatites are corroborated by the interest areas delimited in previous studies.