Seismic facies identification model based on Swin Transformer
Seismic facies identification is an important technology in the process of oil and gas exploration and development,but it has been facing several problems for a long time.For example,its method model consumes more time in training and prediction and provides interpretation results of strong subjectivity,and the multi-scale features are neglected in the feature extraction of each layer.To deal with low seismic facies identification accuracy and high calculation cost,this paper establishes a Seismic Facies Identification model based on Swin Transformer(SFI-ST).Firstly,the SFI-ST model,combined with the convolutional neural network,captures the detail features of seismic facies continuously using encoder and decoder.Then,the effectiveness of the model is tested and evaluated by using two kinds of data sets.Additionally,considering the influence of data set partition on the model,the performances at different partition proportions are comparatively analyzed.Finally,an ablation experiment and anti-noise analysis are conducted on the model.The following results are obtained.First,the Swin Transformer module used in the encoder has a better capability of feature extraction.The feature extraction strategy based on a small moving window ensures the model to learn the features of high resolution seismic profiles faster,and the application of the self-attention mechanism to compute the features in each moving window ensures the model to extract the local features more accurately under a larger view.Second,in the mode of layer-by-layer feature fusion,the Swin Transformer improves the feature extraction speed while ensuring the model to extract multi-scale features.Third,the fusion of Swin Transformer and convolutional neural network module realizes the feature extraction at each layer,and strengthens the model's capability of extracting outlines and edges.In conclusion,the mean intersection over union of SFI-ST model for the data of two working areas is 73.2%and 77.6%,which is at least 10.7%and 6.0%higher than that of other mainstream deep learning algorithms.And the running time of the SFI-ST model is 0.62 h and 2.88 h,which is at least 15.1%and 24.2%less than that of other mainstream deep learning algorithms.What's more,the SFI-ST model solves the problems of the existing intelligent seismic facies identification method such as low speed and low accuracy to a certain extent,and provides a new method for seismic facies identification,technically assisting the progress of oil and gas exploration and development.
Seismic facies identificationSemantic segmentation modelSwin TransformerMulti-scale featuresOil and gas reservoir prediction
万方数据
维普
