Tips and Tricks for Improving Seismic Facies Classification and Estimating Uncertainty

Abstract:

Seismic facies interpretation provides a reference for analyzing geological conditions and predicting oil and gas reservoirs. In recent years, deep-learning methods, especially pixel-wise semantic segmentation, have been used in seismic interpretation tasks. However, unlike binary classification problems such as fault or salt dome detection, seismic facies classification is a multi-class segmentation task, which has been facing some obstacles. Firstly, as a result of sedimentary tectonics, the spatial distributions of seismic facies are crucial global context information which is effective to produce good quality results, but it is difficult for conventional segmentation models to acquire global-scene-level prior of seismic data. Secondly, the labels of seismic facies are often converted into one-hot encoding when we train a multi-class classification model, which is easy to cause overfitting and overconfidence due to noisy labels. Finally, uncertainty estimation is important for a prediction model because it can make us know the “confidence” with which we can trust our prediction for decision making. But it is hard for multi-class segmentation models to visualize probability maps. Given these issues, we implement our facies classification network by simplified U-Net with pyramid pooling module which empirically proved to be an effective global contextual prior. Then, we make a prior label and then iterate through several successions of training a new model by using the previously trained model as a Label Refiner, so the subsequent network models are trained with more accurate labels. Eventually, given the model with less overconfidence, we present an approach for estimating the uncertainty of classification results by calculating Prediction Entropy. By training diversity samples after special data augmentation and tuning parameters, the application of our scheme on F3 dataset demonstrates its improvements and we observe high prediction uncertainty at facies boundaries.