Characterization of fault-karst reservoirs based on deep learning and attribute fusion

The identification of fault-karst reservoir is crucial for the exploration and development of fault-controlled oil and gas reservoirs. Traditional methods primarily rely on well logging and seismic attribute analysis for karst cave identification. However, these methods often lack the resolution needed to meet practical demands. Deep learning methods offer promising solutions by effectively overcoming the complex response characteristics of seismic wave fields, owing to their high learning capabilities. Therefore, this research proposes a method for fault-karst reservoir identification. Initially, a comparative analysis between the improved U-Net++ network and traditional deep convolutional networks is conducted to select appropriate training parameters for separate training of karst caves and faults. Subsequently, the trained models are applied to actual seismic data to predict karst caves and faults within the research area, followed by attribute fusion to acquire data on fault-karst reservoirs. The results indicate that: (1) The proposed method effectively identifies karst caves and faults, outperforming traditional seismic attribute and coherence methods in terms of identification accuracy, and slightly surpassing U-Net and FCN; (2) The fusion of predicted karst caves and faults yields clear delineation of the relationship between top karst caves and bottom fractures within the research area. In summary, the proposed method for fault-karst reservoirs identification and characterization provides valuable insights for the exploration and development of fault-controlled oil and gas reservoirs in the region.