Research And Implementation On Salt Dome Identification Algorithm Based On Image Segmentation

Salt dome is a geological formation formed by the movement and deposition of underground rock salt layers,and is usually surrounded by a large amount of oil and gas resources.Accurately identifying the distribution and morphology of salt dome is of great significance for oil and gas exploration and development.The traditional approach to salt dome identification relies on seismic attributes and geological experts’ theoretical knowledge and personal experience,which is inefficient and subjective.In recent years,deep learning technology has developed rapidly and been introduced into the application of seismic data interpretation.Compared with the traditional manual interpretation methods,deep learning methods have higher automation and better robustness.This paper use deep learning technology to transform the salt dome identification problem into a segmentation problem for processing seismic images,introduce the image semantic segmentation algorithm into the salt dome identification task,and improve the deficiencies of existing image segmentation algorithms according to the characteristics of seismic images,and finally verify the validity of the algorithms in this paper through vertical and horizontal comparisons.The main contents of this paper are as follows:(1)A salt dome identification algorithm based on SE-Unet is proposed.The salt dome identification problem is transformed into a bipartition problem of identifying salt dome features by labeling salt dome and backgrounds on seismic images.Based on the structure of the conventional U-Net network,the salt dome features are deeply mined by combining SENet154 as the backbone network of the encoder.According to the characteristics of salt dome in seismic image,the Lovasz-Softmax loss function,which is more suitable for small object segmentation,is selected to improve the segmentation effect of salt dome boundary and effectively complete the task of salt dome identification in seismic image.Through experiments on the dataset provided by the TGS Salt Body Identification Challenge,the results show that SE-Unet have the best performance compared to other salt dome identification algorithms,with the segmentation accuracy of 97.5% and the Io U of 87.26%.(2)A salt dome identification algorithm based on Deep Labv3+ is proposed.Firstly,the Deep Labv3+ network is used to segment salt domes.Then,to address issues such as insufficient accuracy in segmentation results and inaccurate boundary detail segmentation,the Deep Labv3+model is improved to further improve the accuracy of salt dome identification in seismic images.The attention mechanism is introduced into the Deep Labv3+ network to enhance different class features between classes and different local features within classes,and the improved Deep Labv3+ salt dome identification algorithm is proposed.The effectiveness of the module is demonstrated by ablation experiments,and compared with the Unet-based method,PSPNet,USKNet,Se Net154-FPN,SE-Unet and Deep Labv3+ algorithms,the results show that the improved Deep Labv3+ algorithm performs optimally in the accuracy and the Io U,reaching 98.4%and 90.68%,reflecting the superiority and applicability of the algorithm.The algorithm proposed in this paper has been proved through experiments to be able to accurately and efficiently complete the salt dome identification task in original seismic images,which has high research significance and practical value.