Bottom Simulating Reflectors Identification Based On Convolutional Neural Network

Natural gas hydrate is a clean energy source in the 21 st century.Under the guidance of carbon neutrality policy,hydrate geophysical exploration will become a research hotspot in the future.The main purpose of geophysical exploration is to find underground target bodies,and after collecting and processing geophysical data,inversion or migration can obtain relevant parameters of target bodies.In hydrate seismic exploration work,the characteristics of seismic images are not always obvious,and the existing hydrate exploration data processing and interpretation methods have fallen behind the actual needs.In recent years,with the development of computers,machine learning has been widely used in various disciplines,which transforms an abstract classification problem into a nonlinear mapping problem by iteratively updating the weight parameters.In this paper,the neural network in deep learning is applied to the polarity recognition of bottom simulating reflector(BSR)of the natural gas hydrate,and the geophysical characteristics and image recognition methods of hydrate are systematically introduced.The network performance is optimized to improve the recognition accuracy of the network for BSR.The main research work is as follows:(1)This paper derives the decision parameters of the Softmax loss function in the convolutional neural network for class discrimination,and applies the Policy Gradient to the convolutional neural network loss function.In the process of parameter search,so as to achieve accurate identification,this method is suitable for the classification network using Softmax loss function as the classifier.(2)This paper generates training samples by simulating BSR profile images,and applies the improved network based on the classical convolutional neural network to the simulated profile classification work,which proves that the improved network has higher performance in the BSR polarity discrimination task.Accuracy,and the training samples are windowed in this paper,which provides a new perspective for the automatic and intelligent research of geophysical image recognition with insufficient data.(3)In this paper,the real seismic section is applied to the improved neural network,and the visual interpretation method is used to highlight the BSR features.The network also shows the highest accuracy rate in the real data,and the effectiveness of the method is proved according to the visual feature interpretation.