[S-TT34] Airborne surveys and monitoring of the Earth

[STT34-P06] Application of Deep Learning Inversion for Frequency-Domain Airborne Electromagnetic Field Data

*Seiya Oue¹, Takumi Ueda¹, Shigeo Okuma², Toshiyuki Yokota² (1.WASEDA Univercity, 2.National Institute of Advanced Industrial Science and Technology)

Keywords:Electromagnetic Exploration, Airborne Electromagnetic Exploration, Deep Learning, Inversion

The airborne electromagnetic method has the advantage of being able to survey broad areas efficiently. However, the number of survey points and data becomes enormous, making the cost of inversion very high. On the other hand, deep learning inversion as a regression problem is considered capable of rapidly inverting large-scale data such as airborne electromagnetic surveys by utilizing a trained deep neural network model. Therefore, in this study, we applied deep learning to 1D inversion of frequency domain airborne electromagnetic field data measured by the RESOLVE system. Specifically, we applied the deep learning code under development to the RESOLVE field data measured in 2012 along the Pacific coast of the Tohoku region after the tsunami from the Great East Japan Earthquake inundated the area.

As a result, it was demonstrated that the setting range of resistivity in the training data greatly affected the inversion results. The RMSPE (Root Mean Squared Percentage Error) value, which is a measure of the residual between the observed data and predicted data from forward modeling of the estimated resistivity structure by deep learning, showed the smallest values when the resistivity generation range was set to 0.1 to 1000 Ohm-m. The results were generally consistent with other inversions based on the gradient method.

Future work includes validation of the hyperparameters of deep learning fixed in this study, generation of training data that accounts for electromagnetic noise, improvement of the network model, and application of this method to various field data.