基于LinkNet的地震相自动划分

Automatic seismic facies classification based on LinkNet

地震相人工解释需要耗费大量的时间和精力,并存在很大的主观性和不确定性,从而直接影响地震资料解释的准确性。虽然深度学习算法已经广泛用于地震相划分,然而由于地震相出现的模式及其空间尺度的多样性,在保证高分辨率以及高精度的同时,提高计算效率仍是一项具有挑战性的任务。为此,提出基于LinkNet的地震相自动划分方法,采用多分类交叉熵与Tversky的加权线性组合作为网络训练的损失函数。Tversky通过调整参数平衡假正类和假负类,进而提升召回率等指标以提高不均衡数据中少数类地震相边界的刻画精度。LinkNet解码层共享编码层的学习特征,使解码层的结构更精简,大大提高了计算效率。在荷兰北海F3区块的测试结果表明:所提方法刻画地震相的精度高于U-Net+PPM(金字塔池化模块),在面对不均衡数据时,对占比较小的类别的关注度更高,并具有更好的边界刻画能力;LinkNet计算速度快,可以在配置更低的设备上运行,较U-Net+PPM更实用。

The artificial interpretation of seismic facies requires a lot of time and energy and is highly subjective and uncertain,which directly affects the accuracy of seismic data interpretation.Although deep learning algorithms have been widely used in seismic facies classification,due to the diversity of patterns and spatial scales of seismic facies,it remains a challenging task to improve computational efficiency while ensuring high resolution and precision.Therefore,an automatic seismic facies classification method based on LinkNet is proposed.The weighted linear combination of the multi-class cross entropy and Tversky is used as the loss function of network training.Tversky improves the description accuracy of the boundaries of minority classes of seismic facies in the unbalanced data by adjusting the parameters to balance the false positive and false negative classes and improving the recall rate and other indicators.LinkNet's decoding layer shares the learning characteristics of the encoding layer,which makes its structure simpler and greatly raises computational efficiency.The tests on the F3block of the North Sea in the Netherlands show that the accuracy of the proposed method for describing seismic facies is higher than that of U-Net+PPM,and when faced with unbalanced data,it pays more attention to the minority classes and has better boundary characterization ability.LinkNet has fast computing speed and can run on devices with lower configurations,which is more practical than U-Net+PPM.