3D density inversion of gravity anomalies based on UNet plus

3D density inversion is a hot research topic in geophysics. It is more important to inverse quickly and efficiently using gravity data in the context of big data and artificial intelligence development. In traditional inversion methods, it takes up a large memory and a long time to store the large coefficient matrix. At the same time, it is difficult to determine the parameters of the regularization constraint term added to the constraints of the inversion results. While deep learning does not rely on a priori information, nor does it need to compute and store coefficient matrices, which makes the computation much more efficient. Based on this, this paper proposes a gravity anomalies inversion method using UNet+ network. Firstly, we changed some parameters of UNet+ network. LeakyReLU, which has a more stable gradient when the absolute value of input data are large, is selected as the activation function. Batch Normalization layer is added to enhance the convergence speed and stability of the network. Then, in order to improve the global optimization capability of the network, a learning rate updating strategy based on cosine annealing is used. The network is trained on the data sets and the label sets by Adam optimization algorithm using first-order and second-order moment estimation of gradients to achieve the 3D density inversion of gravity anomalies. The fast and stable convergence ability of UNet++ and LeakyReLU and the stronger global optimization finding ability of the cosine annealing are verified. The experiments of the noise-containing models and practical data inversion results further prove the correctness, effectiveness, good generalization and noise immunity of the method.