Research On 3D Geological Modeling Based On Video Super-resolution And Reliability Evaluation Method

3D geological modeling is an important part of stratigraphic research in geological exploration,petroleum engineering and so on.Traditional geological modeling requires a lot of manual intervention,especially in key steps such as drawing stratigraphic profiles which are often cumbersome,complex and difficult to modify processes.In recent years,with the rapid development of deep learning technology,especially its end-to-end mechanism and excellent representation learning ability,provides new ideas for solving the problems of high labor cost,difficult data acquisition and complex training methods in traditional technology.Aiming at the problems of data sparsity and process complexity in 3D geological modeling,from the perspective of computer vision and image processing,this thesis takes the stratigraphic profile that generated from the original stratigraphic data as the input data set,and takes the 3D geological modeling method and its reliability evaluation strategy based on video super resolution technology.The main task of the geological modeling method is to complete the generation,supplementation and refinement of stratigraphic profile data sets in the space and time domain.The work of this thesis consists of three parts:(1)A 3D geological modeling method based on deep learning video super-resolution technology is designed.The method in this thesis draws on the basic structure of optical flow generation to intermediate frame synthesis.The network structure mainly includes two parts.The first part is the model of extracting the geological change field based on 3D deformable and pixels-adaptive convolution.And the new concept "geological change field" is used to represent the formation characteristic flow,using 3D convolution that can be adaptively adjusted according to features such as pixel content and object shape to extract the geological change field;The second is the spatiotemporal super-resolution prediction network based on the 3D convolution.This sub-module is based on the geological change field generated by the first sub-module,and integrates the calculated warping image and the feature flow information of the intermediate moment to complete the generation of the intermediate frame;(2)The loss measurement strategy according to the particularity of geological image content information is optimized.In the prediction of stratigraphic profile,there are often problems such as fuzzy generated map and inaccurate prediction of stratigraphic boundary.Accordingly,this thesis proposes a training mode based on deep learning discriminator.By analyzing the experimental results of the basic module,an additional deep learning edge detection network is trained as a discriminator.The loss calculation method of the model is further improved by it;(3)A comprehensive evaluation system for the reliability of the 3D geological modeling method is proposed.Due to the intersection of the work content,on the one hand,the reliability of the 3D geological modeling method is evaluated from the perspective of traditional geology;On the other hand,the thesis uses the method of computer vision to objectively evaluate the generated images,and finally visualize it.The comparison experiments prove that the super-resolution 3D geological modeling method proposed in this thesis can effectively reduce the modeling complexity and improve the model accuracy for the data sparse problem.Especially,when dealing with complex strata,it can greatly reduce the stratum boundary error.In addition,the method is also competent for general video super-resolution tasks in the field of computer vision,such as video repairing.