CapsNet-Enhanced Seismic Event Classification: Benchmarking and System Deployment on the DiTing 2.0 Dataset

Summary Automated seismic event classification is a critical component of modern earthquake monitoring and network operations, yet many previous studies have been limited by small datasets or binary tasks. This work presents the first systematic benchmark of mainstream deep learning models, including convolutional neural networks (CNNs), Transformer-based architectures, and Capsule Networks (CapsNet), on the large-scale DiTing 2.0 AI seismic dataset, which contains 19,384 labeled three-component waveforms spanning three classes (natural earthquakes, quarry blasts, and mine collapses). Thus, we provide reproducible baselines for the seismic events classifications. Several CapsNet variants optimized for DiTing 2.0 are evaluated, and our results show that the CapsNet+Res model using MFCC representations and data augmentation achieves 91.08% accuracy (weighted F1 = 91.10%) on the held-out test set. The multi-station voting further improves event-level accuracy to 97.52%, while a companion noise-event classifier attains 98.47% accuracy. Functionality testing on demonstration continuous records confirms reliable end-to-end operation within a transferable and user-friendly system, underscoring its feasibility for seismic network applications. Overall, this study bridges methodological development and operational application by providing robust baselines, insights into the interplay of input features and architectures, and a practical platform for automated classification; large-scale continuous-data evaluation, cross-regional transfer validation, and adaptive learning strategies remain important directions for future research.