Automatic assessing and masking algorithms for electromagnetic transfer functions based on machine learning methods

Abstract

The magnetotelluric (MT) method is an electromagnetic geophysical technique to investigate the Earth's electrical conductivity structure. It has been widely applied from deep structural studies to near-surface resource explorations. The raw MT time series data are usually transformed and processed into frequency-domain transfer functions (TFs), before being applied in geophysical inversion and interpretations. However, the assessment and elimination of low-quality TF data points still rely heavily on manual operations, which are time-consuming and requiring substantial expertise for operators, thereby reducing the overall efficiency of the data processing workflow. The rise of machine learning (ML) algorithms nowadays has opened up possibilities for rapid and automated data classification, leading to extensive success in fields like finance and image processing. This study explores two popular ML classification algorithms, namely Support Vector Machine (SVM) and Deep Neural Network (DNN), to automatically assess the TF quality. Various data features, such as the difference between a given data point and its neighboring points, were extracted to form a reduced subspace for classification. The classification accuracy of the two algorithms was compared against their manual counterpart, which indicates that the SVM algorithm achieved an accuracy of 93 %, while the DNN algorithm achieved 86 % with the real-world TF data. Consequently, an automated electromagnetic TF masking program based on the SVM algorithm was developed, enabling the accurate and rapid identification and removal of low-quality data points. For instance, manual masking of TF data from a single site may typically require approximately five minutes, whereas the new algorithm accomplishes the task in just a few seconds, significantly enhancing the efficiency of data masking.