Transfer learning-based hybrid deep learning method for gas-bearing distribution prediction with insufficient training samples and uncertainty analysis

The availability of sample data, required for seismicity-based reservoir prediction using deep learning (DL), is often limited by well availability. To alleviate this problem, this study proposes a hybrid DL method combining unsupervised learning, transfer learning (TL), and intelligent optimization algorithms to predict gas-bearing distributions. First, three unsupervised learning methods were used to optimize multicomponent seismic attribute data during preprocessing and improve the sample data quality. Subsequently, a convolutional neural network (CNN) model was pre-trained using synthetic data, and the adaptive mutation particle swarm optimization (AMPSO) algorithm was used to optimize the hyperparameters of the CNN model during training to obtain the AMPSO-CNN model. Finally, the partial convolutional layer parameters of the pre-trained network were transferred, and the remaining parameters were adjusted using real data samples to obtain the TL-AMPSO-CNN model. The trained model was used for gas-bearing distribution prediction, and its stability was assessed using uncertainty analysis. The results confirmed that the hybrid DL model performed efficiently in gas-bearing prediction compared to other individual models. Thus, for datasets with insufficient training samples, the proposed scheme can be applied as a feasible alternative to predict gas-bearing distribution.