Application of geophysical and multispectral imagery data for predictive mapping of a complex geo‑tectonic unit: a case study of the East Vardar Ophiolite Zone, North‑Macedonia

Abstract

The Random Forest (RF) and K nearest neighbors (KNN) machine learning (ML) algorithms were evaluated for their ability to predict ophiolite occurrences, in the East Vardar Zone (EVZ) of central North Macedonia. A predictive map of the investigated area was created using three data sources: geophysical data (digital elevation model, gravity and geomagnetic), multispectral optical satellite images (Landsat 7 ETM + and their derivatives), and geological data (distance to fault map and ophiolite outcrops map). The research included a comparison and discussion on the statistical and geological findings derived from different training dataset class ratios in relation to a testing dataset characterized by significant class imbalance. The results suggest that the precise selection of a suitable class balance for the training dataset is a critical factor in achieving accurate ophiolite prediction with RF and KNN algorithms. The analysis of feature importance revealed that the Bouguer gravity anomaly map, total intensity of the Earth’s magnetic field reduced to the pole map, distance to fault map, band ratio BR3 map obtained from multispectral satellite images, and digital elevation model are the most significant features for predicting ophiolites within the EVZ. KNN showed poorer results compared to RF in terms of both the evaluation metrics and visual analysis of prediction maps. The methods applied in this research can be applied for predictive mapping of complex geo-tectonic units covered by dense vegetation, and may indicate the presence of these units even if they were not previously mapped, particularly when geophysical data are used as features.