Machine Learning Methods for Development of a Model of “Land/Ocean” Contrast in the Parameters of ULF/ELF Variations of the Geomagnetic Field

The paper is devoted to verification of a hypothesis about existence of a “land/ocean” contrast in the parameters of ULF/ ELF variations of the geomagnetic field in the upper ionosphere. To solve this problem, a model is developed using machine learning methods. Modeling is based on magnetometer data recorded by an ionospheric satellite SWARM. The optimal shape of the feature space is constructed using the principal component method. The model is built using ensemble method of gradient boosting on decision trees. The quality of model is evaluated using classification metrics and curves of threshold values of decision function in metrics spaces. Interpretation of modeling results is based on the method of permutation importances. The results of model evaluation show that the model combines statistically stability and high accuracy. Thus, land-ocean contrast really reveals itself in parameters of natural electromagnetic fluctuations in ULF -ELF band in the upper ionosphere.