A Machine Learning-Based Nonlinear Regression Application for Processing Geomagnetic Data Reconstruction

Abstract

When it comes to near-surface exploration, finding unexploded ordnance, and other applications that rely on geomagnetic data, the accuracy and reliability of such data are key considerations. Based on machine learning methods, this research proposes a geomagnetic data reconstruction method for undersampled geomagnetic data. When compared to the conventional linear interpolation methods, the suggested methodology is more time efficient and lower in labour cost. The support vector machine, random forests, and gradient boosting models were all developed in this study. Recurrent neural network, a deep learning method, was also used to boost training performance. A continuous regression hyperplane was generated from a training dataset using the suggested learning methods. Using the provided regression hyperplane, the relationship between the mock-up missing data and the rest of the data is mapped out. Finally, the hyperplanes that were trained were utilised to rebuild the missing geomagnetic traces for validation, and they may be used to reconstruct further gathered fresh field data. In the end, numerical experiments were developed. When compared to the standard linear technique, our methods produced better results, with a reconstruction accuracy that was improved by 10% to 20%.