The Denoising of Seismic Signals Based on Joint Loss and Feature Fusion Network

To suppress noise of varying levels in seismic signals and obtain effective seismic data, a denoising approach based on joint loss and feature fusion network is proposed. The network architecture comprises two parts: a 5-layer convolutional neural network (CNN) and an improved U-Net network. The 5-layer CNN preprocesses noisy data by estimating noise in advance to mitigate interference from prior knowledge. Subsequently, a feature fusion technique is employed to combine original seismic data with the processed data from the 5-layer CNN, which is then fed into the second part of the network for denoising. The enhanced U-Net network structure incorporates convolutional block attention module (CBAM) and residual structures in the encoder, utilizes bilinear interpolation in place of deconvolution in the decoder, and employs leaky rectified linear unit activation function. Jointly utilizing mean square error (MSE) and mean absolute error (MAE) as loss functions aims to optimize denoising performance. Experimental validation using publicly available synthetic and real seismic data demonstrates the effectiveness and robustness of the proposed algorithm. Results indicate that under various levels of Gaussian white noise, the proposed method outperforms DnCNN and U-Net algorithms in achieving higher signal-to-noise ratios and better preservation of coaxial texture information.