Optimizing noise reduction in layered-earth magnetotelluric data for generating smooth models with artificial neural networks

Abstract

The presence of noise in geophysical data poses significant challenges to accurate analysis and interpretation, impacting the reliability of geoscience research and exploration. In the context of inverting electromagnetic-sounding data, the objective is to derive a unique model for interpreting observations while acknowledging the non-uniqueness of solutions. The uncertainty introduced by unwanted signals complicates the selection of an initial model for inversion. This study emphasizes the heightened efficacy of the artificial neural network (ANN) model, prioritizing smoothness to mitigate overinterpretation and eliminate arbitrary discontinuities in layered models. The goal is to identify the smoothest model fitting experimental data within a defined tolerance rather than maximizing model roughness. A practical ANN scheme is developed, predicting subsequent values based on previous ones while optimizing for arbitrary discontinuities. Extensive evaluation using synthetic and real-world magnetotelluric data showcases the model's performance. Employing a sliding window technique for dataset preparation allows the extraction of local patterns and trends in time series data. The results demonstrate the remarkable noise reduction capabilities of neural networks, surpassing traditional methods like filtering and wavelet transform. The neural network model consistently produces predictions with relatively low Mean Absolute Error (MAE) values, indicating its ability to preserve underlying geological structures even in noisy conditions. Specifically, MAE for actual inverted data ranges from 0.49 to 5.95, while MAE for predicted values by the neural network model ranges from 6.19 to 7.75. Notably, the model outperforms wavelet transform, particularly in preserving short trends during noise reduction. This aligns with prior studies emphasizing neural networks' superior performance in handling complex data patterns. Further exploration applies the neural network model, revealing accuracy rates of approximately 93% along the east–west (EW) direction and 92.5% along the north–south (NS) direction for six diverse profiles. Robustness is demonstrated by introducing various noises into testing sample data, showcasing the model's resilience in inversion findings. This study underscores the profound effectiveness of neural networks in noise reduction, highlighting machine learning's vast potential in geophysical data analysis. Beyond conventional techniques, these insights offer valuable implications for the future of geophysics research and applications.