Abstract: In order to reduce the influence of interference noise on gravity measured data and further improve the accuracy of gravity data processing, a gravity data denoising method based on Multilevel Wavelet Residual Network (MWRNet) was proposed, which combined the wavelet transform and neural network to realize the removal of noise components in gravity data. Firstly, the gravity data were decomposed by wavelet transform, and then a neural network was utilized for noise extraction, while the Residual Channel Attention (RCA) module was introduced to enhance the noise extraction ability of the network. The proposed gravity data denoising method was tested using simulated data and measured data, and the experimental results show that the proposed method has better results compared with other gravity data denoising methods. In an experiment with a noise standard deviation of 50, the proposed method has a 21% and 9% improvement in Peak Signal-to-Noise Ratio (PSNR) and Structure Similarity Index Measure (SSIM), respectively, compared with traditional denoising algorithms Block Matching 3D (BM3D), and more than 0.4% and 1% improvement compared to the deep learning-based denoising algorithms Denoising Convolutional Neural Network (DnCNN) and Multi-level Wavelet-CNN (MWCNN), respectively.

Key words: gravity data, wavelet transform, deep learning, residual connectivity, channel attention

摘要： 为降低干扰噪声对重力实测数据的影响，进一步提高重力数据处理精度，提出一种基于多级小波残差网络(MWRNet)的重力数据去噪方法，该方法结合小波变换和神经网络实现对重力数据中噪声分量的去除。首先通过小波变换对重力数据进行分解，然后利用神经网络进行噪声提取，同时引入残差通道注意力（RCA）模块增强网络的噪声提取能力。利用模拟数据和实测数据对所提重力数据去噪方法进行测试，实验结果表明：所提方法相较于其他重力数据去噪方法有着更好的效果。在噪声标准差为50的实验中，所提方法相较于传统去噪算法三维块匹配算法（BM3D）在峰值信噪比（PSNR）和结构相似性指数（SSIM）上分别有着21%和9%以上的提升，相较于基于深度学习的去噪算法DnCNN（Denoising Convolutional Neural Network）和MWCNN（Multi-level Wavelet-CNN）分别有着0.4%和1%以上的提升。

关键词: 重力数据, 小波变换, 深度学习, 残差连接, 通道注意力