Automatic seismic feature recognition using deep convolutional neural networks

The accurate and efficient identification of seismic signal signatures is a fundamental requirement for the development of effective seismic detection and early warning systems. In this study, we propose an advanced Deep Convolutional Neural Network (DCNN) architecture, which is designed to automatically identify and classify seismic features with high accuracy. Our model builds upon existing ResNet and Inception models, integrating a multi-scale feature extraction layer and employing residual connections to enhance the network depth without affecting training efficiency. Furthermore, the attention mechanism and a hybrid loss function combining cross-entropy and focal loss are incorporated into the model to address the issue of category imbalance and enhance the capacity for feature discrimination. To enhance the model's complexity and robustness, we employed ensemble learning techniques and utilised a pre-trained network on a large-scale dataset for transfer learning. These innovations facilitate the network's capacity to discern intricate patterns within seismic data, a capability that is markedly superior to that of traditional methods, and to achieve more accurate feature recognition. The experimental results demonstrate that our DCNN model outperforms existing methods in terms of recognition accuracy, thereby validating its efficacy in real-time seismic monitoring and analysis applications.