PERFORMANCE ANALYSIS OF DEEP NEURAL NETWORK FOR AUTOMATIC SEISMIC PHASE PICKING IN THE SUMATRA REGION

• Main Author: Wahyu Saputra, Rifky
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/84759
• Institution: Institut Teknologi Bandung

A more efficient and automated approach to earthquake data analysis is crucial for reducing intensive manual intervention in determining seismic wave phases. This study focuses on the development and performance analysis of automatic seismic wave phase determination using a deep neural network approach, specifically in the Sumatra region of Indonesia. The data used in this study were obtained from the Meteorology, Climatology, and Geophysics Agency (BMKG) for the years 2009?2020, covering 131 recording stations that documented 1356 seismic events with magnitudes greater than 1.8. Three types of models based on the PhaseNet architecture were utilized: a pre-trained model, the retrained Sumatra model, and a transfer learning model. The pre-trained model demonstrated varying performance with F1-Scores ranging from 0.095 to 0.717 for the P-wave phase and 0.033 to 0.485 for the S-wave phase. Subsequently, the model was retrained with ±6000 basic waveform data which was then augmented to train the Sumatran model, which improved the F1-Score to 0.728 for the P-wave phase and 0.388 for the S-wave phase. Further experimentation involved modifying the PhaseNet architecture by incorporating attention gates, LSTM, Bi-LSTM, and self-attention layers. These modifications further enhanced the F1-Score to 0.749 for the P-wave phase and 0.454 for the S-wave phase. The application of transfer learning through fine-tuning or freezing layers of the pre-trained model yielded better F1-Scores, ranging from 0.709 to 0.774 for the P-wave phase and 0.488 to 0.599 for the Swave phase. From these results, it is concluded that transfer learning is more advantageous in terms of time, data volume, and performance. Full model training can be conducted if sufficient and varied training data are available, allowing the model to learn more accurately. Further development is needed by increasing the amount of data, utilizing different model architectures, and further modifying the PhaseNet architecture.