Pseudo-2D deep learning inversion of frequency-domain airborne electromagnetic data considering terrain effects

Abstract

Pseudo-2D deep learning inversion of frequency-domain airborne electromagnetic data considering terrain effects Minkyu Bang Dept. of Earth Resources and Environmental Eng. The Graduate School Hanyang University Deep neural networks (DNNs) have recently been used to interpret frequency domain electromagnetic (EM) data and they offer ways to interpret the vast amounts of data rapidly. In particular, the application of DNN-based inversion to the airborne EM data is highly anticipated because airborne survey acquires vast amount of data. However, since airborne surveys are typically conducted in mountainous regions, severe terrain changes can distort the EM data and lead to unreasonable interpretations without correction for the terrain effects. In addition, DNN-based inversions that reduce the terrain effects have not yet been proposed, while some conventional inversion techniques can correct distortions due to the terrain effects. Since the performance of the DNN depends on the amount and characteristics of the training dataset, it is important to generate various pairs of airborne EM (AEM) data and resistivity model assuming various subsurface situations. However, it requires a lot of variables to construct two-dimensional (2D) or three-dimensional (3D) resistivity models with diverse terrain patterns, and it takes a lot of time and computational cost to calculate the EM responses for each corresponding resistivity model. Therefore, this thesis proposed the DNN-based pseudo-2D inversion method, which can consider the terrain changes using the EM responses and terrain information obtained from three adjacent stations. The terrain information is converted into the slope angle of the left and right stations based on the center of the three stations and it is used for the input of DNN training. To evaluate the performance of the trained DNN, several numerical experiments were conducted. In addition, in order to confirm the importance of the terrain information in the DNN training process, the results of numerical experiments of the trained DNN were compared with those of the one-dimensional (1D) Gauss-Newton inversion and those of the DNN trained without terrain information. As a result of numerical experiments, the DNN trained with terrain information successfully mitigated the distortions of EM data caused by the terrain effects, and showed more reasonable results than other methods. Furthermore, the trained DNN’s field applicability was evaluated by applying it to a field AEM dataset acquired in Southcentral Alaska. The trained DNN provided reasonable results even though the data were obtained in mountainous terrain with severe terrain changes. The proposed method is anticipated to enhance the field applicability of AEM survey because it can interpret the AEM data very quickly without any hyperparameter adjustment using the trained DNN regardless of the terrain changes. Keywords: Deep neural networks (DNNs), frequency domain airborne electromagnetic (AEM), pseudo-2D method, terrain effects