| In seismic exploration,the seismic records acquired and observed are pre-stack data,which does not provide direct information on the subsurface structure.In order to obtain more accurate information about the subsurface structure,the data is stacked after preprocessing and arranged in a certain position to form the post-stack seismic data.Compared with pre-stack data,post-stack data can provide clearer and more accurate information about the subsurface structure and contains a wealth of information about the geological units.However,traditional post-stack seismic data interpretation methods are highly multi-solution and often require a combination of geophysical and geological knowledge and experience.This process is time-consuming and labour-intensive,and is highly influenced by subjective human factors.Post-stack seismic data,as an amplitude section,that visually reflects the structure of the subsurface,have some similarity with the single-channel grey-scale image.Therefore,in order to reduce the time cost associated with manual seismic interpretation,it is necessary to introduce deep learning methods with strong feature learning capability and high generalisation into the interpretation of post-stack seismic data.With this objective in mind,this thesis investigates the application of deep learning methods to various post-stack seismic data interpretation problems.The construction of network models,the design of objective functions,and the evaluation and uncertainty analysis of prediction results are investigated theoretically.The research is based on deep learning methods for post-stack seismic data reconstruction,and research on the most difficult tasks in seismic interpretation: multi-classification and small-scale geological target identification problems.This research has theoretical significance and practical value for geophysical interdisciplinary research.It is also important in exploration and development of complex oil and gas reservoirs.The main specific work is as follows:(1)Reconstruction of post-stack seismic data based on generative adversarial networks: The impediments to the use of post-stack seismic data for target identification and interpretation are analysed.A generative adversarial network is used to achieve both post-stack seismic data interpolation and super-resolution.The structure of the generative adversarial network model,the loss function and the training data production for the reconstruction of post-stack seismic data are investigated.(2)Research on multi-classification tasks for post-stack seismic data based on deep learning: The seismic facies classification based on deep learning method is improved,and the application of deep learning multi-classification tasks to post-stack seismic data target identification problems is analysed.Deep learning models,loss functions,label structures and uncertainty assessment for seismic facies classification are investigated.(3)Research on small-scale geological target recognition based on deep learning:Taking seismic cave data as an example,two deep learning models are used to achieve seismic anomaly recognition and geological anomaly prediction respectively,where the input of geological anomaly prediction is the output of seismic anomaly.The model structure and loss function for the small-scale geological target recognition network are investigated.The main innovations in this thesis are:(1)The two tasks of interpolating and improving the resolution of post-stack seismic data are combined,and a generative adversarial network is used to achieve simultaneous interpolation and super-resolution of post-stack seismic data.In the seismic data reconstruction problem,a loss function based on perceptual similarity is used instead of the traditional objective function that calculates the pixel-by-pixel error.(2)The models and loss functions of seismic facies classification by deep learning has been improved.A new seismic facies classification model and loss function are designed.And a dynamic generation method of seismic phase labels based on label refinement is introduced to invent an a priori label containing seismic and geological information instead of the conventional one-hot label.(3)Proposing a new metric for uncertainty evaluation: Prediction Information Entropy.It can be used for uncertainty analysis and visual presentation of deep learning multi-classification seismic target classification results.(4)A "two-step approach" is designed for small-scale geological targets recognition.The deep learning method is used to identify seismic anomalies and predict geological anomalies respectively. |
