Inversion of Oceanic Parameters Represented by CTD Utilizing Seismic Multi-Attributes Based on Convolutional Neural Network

Abstract

In Recent years, seismic data have been widely used in seismic oceanography for the inversion of oceanic parameters represented by conductivity temperature depth (CTD). Using this technique, researchers can identify the water structure with high horizontal resolution, which compensates for the deficiencies of CTD data. However, conventional inversion methods are model-driven, such as constrained sparse spike inversion (CSSI) and full waveform inversion (FWI), and typically require prior deterministic mapping operators. In this paper, we propose a novel inversion method based on a convolutional neural network (CNN), which is purely data-driven. To solve the problem of multiple solutions, we use stepwise regression to select the optimal attributes and their combination and take two-dimensional images of the selected attributes as input data. To prevent vanishing gradients, we use the rectified linear unit (ReLU) function as the activation function of the hidden layer. Moreover, the Adam and mini-batch algorithms are combined to improve stability and efficiency. The inversion results of field data indicate that the proposed method is a robust tool for accurately predicting oceanic parameters.