基于改进深度学习方法的地震相智能识别

基于深度学习方法的地震相智能识别技术可以大幅度减少人工操作。现有深度学习方法的网络模型只能提取单一接收域下的目标特征，难以获取地震相在剖面上的全局空间分布信息，模型对少数类地震相的边界刻画效果较差，且缺乏对预测结果可靠程度进行评估的手段。针对这些问题，提出一种用于地震相分类识别的深度学习方法：在U-Net模型的末端加入金字塔池化模块以提高模型获取全局信息的能力；采用一种融合交叉熵与Dice指数的目标函数，改善不均衡数据中少数类地震相边界的刻画问题；提出“预测信息熵”的概念用于评估地震相预测结果的不确定性。该研究方法应用于F3工区地震相预测的实验结果表明：改进深度学习方法在地震相预测中具有更高的精度和更良好的边界刻画能力；同时，预测信息熵指标也能够较好地评价预测结果的不确定性。

Intelligent seismic facies classification based on an improved deep learning method

Intelligent seismic facies classification based on deep learning can greatly reduce manual operations.However,when using conventional deep learning methods for seismic facies recognition,the network model can only extract the feature map on a single receptive field,and it is difficult to obtain the global spatial distribution on seismic sections.In addition,the prediction of the boundary of minor seismic facies is inaccurate,and there is not a method for assessing the reliability of prediction on multi-class segmentation models.We propose a facies classification network by simplified U-Net with Pyramid Pooling Module which has been empirically proved to be an effective global contextual prior.And an objective function combining cross-entropy and Dice loss is adopted to improve the boundary characterization of minor seismic facies in unbalanced data.We present Prediction Entropy for estimating the uncertainty of classification results.Applied to F3 dataset,the improved method can enhance prediction accuracy and boundary characterization,and the index of Prediction Entropy can evaluate the uncertainty of the prediction results.