Cave Identification And Quantitatively Determination Via Deep Learning Method

China's carbonate rocks are rich in oil and gas resources,but carbonate reservoirs have complex characteristics such as deep burial,high temperature and high pressure,high-yield,supercritical and swelled and developed caverns.Large-scale fracture-cavity oil and gas fields are composed of a large number of fractures and karst caves.Water,dry or dense oil and gas layers with different oil and gas saturation are distributed between these oil and gas units,so these continuous or discontinuously distributed karst caves are Identifying and quantifying forecasts is the basis for finding relevant oil and gas resources.The cavernous reservoir is generally buried deep,and is usually affected by late diagenesis and tectonic movement.The type is complex,the heterogeneity is strong,the spatial scale of the cave development,the type of filling and the difference of background surrounding rock cause the earthquake of the cave.The reflection characteristics are extraordinarily complex,and the cross-section shows the "bead"-like reflection characteristics with different phase numbers and different reflection energy.In the actual seismic exploration,analyzing the geological structure in a short time,and realizing the identification and quantification of the cave requires the rich experience of the researchers and the knowledge of the earthquake,and the personal subjective method will seriously affect the interpretation results.Therefore,in the face of the analysis of complex geological structures,the more rapid,accurate,efficient and intelligent completion of identification is the goal that geological researchers have been pursuing today.In view of the above problems,this paper studies the characteristics of karst reflection,full convolutional neural network,attribute model fusion,and quantitative calculation of deep regression forest.The specific work is as follows:1.A method for identifying caverns based on multi-model fusion and full convolutional neural networks1)According to the reflection characteristics of karst caves with different scales,shapes and distributions,using forward numerical simulation method to generate different types of representative karst body data,combined with a small number of expertly labeled seismic data,training based on full convolutional nerve The end-to-end pixel-level classification of the network enables efficient identification of karst bodies.2)For the problem that the number of network layers is insufficient due to insufficient sample and the feature extraction is insufficient,the prior knowledge of the seismic field is used to extract the seismic attributes from the seismic amplitude data body as input data to train the full convolutional neural network,and A variety of attribute training models are fused to achieve better karst identification results.2.Proposed quantitative prediction method of karst volume based on deep regression forest1)Using the traditional forward numerical simulation method,based on the quantitative theoretical analysis,the geological body length,height and volume correction coefficient are obtained by engraving and volume estimation statistics of the beaded reflection,so as to quantitatively calculate the cavity volume.2)Using the convolutional neural network to extract the characteristics of seismic data,and using the deep regression forest method to achieve accurate quantitative prediction of the volume of the cave.