MCSEM Weak Signal Denoising Method Based on FARD-NET

The marine controlled-source electromagnetic (MCSEM) method uses low-frequency electromagnetic waves to detect hydrocarbons in the subsurface. However, the presence of noise in the field MCSEM data has a significant effect on the inversion process and data interpretation. To address this issue, we propose a full-scale attention residual double-convolution network model that incorporates multidimensional feature fusion to suppress noise in MCSEM data. Our approach starts with a four-fold downsampling technique that uses residual multiscale double convolution and maximum pooling to extract the nonlinear features of the MCSEM signal at different depths and scales. We enhanced these signal features using an attention module. Furthermore, in the upsampling layer, we combined the features of the MCSEM data from different depths and scales through full-size concatenation. To separate the MCSEM signal from background noise, we implemented a hybrid loss function. The effectiveness of the proposed method is validated by comparing it with existing denoising algorithms, such as wavelet and convolutional neural network algorithms, using both synthetic and field data. The results indicate that, in noise suppression experiments with synthetic data, our method improves the SNR by approximately 19 dB. Furthermore, in noise suppression experiments with field data, time-frequency analyses, and inversion results demonstrate that our method effectively suppresses noise, preserves signal characteristics, and increases the interpretability of the data.