基于深度学习的多属性盐丘自动识别方法

三维地震数据的盐丘解释存在难度大且效率低的问题。为此，以深度学习技术为基础，以少量的二维地震数据为样本训练和测试模型，利用不同地震属性实现盐丘自动识别。流程主要包括三部分：首先，基于盐丘的地震反射特征提取了杂乱、均方根振幅以及方差等3种敏感属性，每一种属性分别选取少量主测线数据及时间切片作为训练样本进行预处理，并利用数据增强方法自动生成大量数据作为网络的训练样本；然后，搭建基于编码—解码器结构的卷积神经网络，分别输入不同属性的两类样本进行模型训练和测试以得到多个独立的模型；最后，为了综合考虑各属性特征，减少预测误差并得到更全面、准确的预测结果，利用集成学习方法融合多个模型并得到优化后的分类结果。结果表明，所提方法能高效、准确地在三维数据体实现盐丘自动分割，盐丘边界清晰，分类错误点明显减少，进一步提高了模型预测能力。

Multi-attribute automatic interpretation of salt domes based on deep learning

It is difficult and inefficient for salt dome interpretation using 3D seismic data.A new workflow uses different seismic attributes to automatically inpterpret salt domes based on a small amount of 2D seismic data as training samples and testing models after deep learning.The wrokflow consists of three parts.First,according to the characteristics of a salt dome on seismic data,extract three types of sensitive attributes including chaotic and RMS amplitude,and variance.For each type of attribute,select a small amount of inline and time slices as training samples and use a data augmentation method to automatically generate massive samples.Second,construct a convolutional neural network based on anencoder-decoder architecture,and input two types of samples with different attributes for training and testing models to obtain multiple independent models.Finally,to comprehensively consider the features of all attributes and obtain more accurate classified results,use an ensemble learning method to merge the models and acquire optimized results.The results indicate that the boundaries of salt domes are clear and the classification errors can be significantly removed.This method can efficiently realize automatic segmentation of salt domes in 3D data setand further improve the prediction ability of models.