A Nonlinear Regression Application Via Machine Learning Methods For Geomagnetic Data Reconstruction Processing

Abstract

Geomagnetic data integrity is a key element in understanding the evolutionary process of the Earth's magnetic field, since it provides useful information for near-surface detection, unexplosive explosive ordnance recognition, etc. This article presents a geomagnetic data rebuilding method based on machine learning techniques aimed at reconstructing geomagnetic undersampled data. The traditional approaches to linear interpolation are prone to time inefficiency and high labour costs, although the proposed method is greatly improved. This paper has been prepared with three classic machine learning models, a vector supporting machine, random forests, and gradient boosting. In addition, a deep learning algorithm, a recurrent neural network, was investigated to further improve training performance. Training data were used to define an ongoing regression hyperplane for the proposed learning models. The hyperplane regression described is a mapping of the connection between the missing data from the mock-up and its intact data. The trained models were then used in hyperplans to recreate the missing geomagnetic traces for validation and to reconstruct additional knowledge collected from the field. Finally, it was derived from numerical experiments. The results showed that, relative to the conventional linear system, the efficiency of our methods was more competitive, as the accuracy of reconstruction was improved by approximately 10% to 20%.