Deep learning is being applied in many aspects of seismic processing and interpretation. Here we look at a deep convolutional neural network (CNN) approach to multiclass seismic facies characterization using well logs and seismic data. In particular, we focus on network performance and hyper-parameter tuning. Significant improvements in predictive capability are possible using data augmentation and automatic and manual tuning. A novel method is introduced to mitigate the effects of class imbalance on network performance, particularly important where the classification problem is characterized by a small amount of labeled data such as availability of well logs. This method involves two training phases. The first phase combines classical training with focal loss categorical cross entropy to partially compensate for class imbalance. The second phase involves retraining after augmenting the under-represented facie classes with samples that are well classified in the first phase (<95% confidence). This is validated on a 3D synthetic lithofacies example based on actual well data and structural information from an onshore oil field.
Presentation Date: Tuesday, September 17, 2019
Session Start Time: 8:30 AM
Presentation Time: 11:25 AM
Location: 221D
Presentation Type: Oral
