TSUNAMI POTENTIAL PREDICTION USING SENSOR FUSION AND MACHINE LEARNING APPROACH FOR EARLY WARNING

• Main Author: Novianty, Astri
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/66548
• Institution: Institut Teknologi Bandung

Tsunamis are seawaves that propagate in all directions as result of an impulsive disturbance on the seabed caused by some changes of the seabed geological structure form in vertical way in a very short time. Most of tsunamis in Indonesia are mainly caused by tectonic earthquake. Not every tectonic earthquake generates tsunami, but most of tsunamis are caused by tectonic earthquake. Indonesia is prone to local field (near field) tsunami disaster because most of its shores are near the source of tsunami. The tsunamis may occur less than 30 minutes after a tectonic earthquake occured. Therefore, tsunami early warning system (TEWS) is very urgently needed particularly in Indonesia. One of important parts in TEWS is decision support system with main task to predict a tsunami event based on a seismic event. When an earthquake occured, it will do some processes to predict whether the occured earthquake is potential to trigger a tsunami or not. There are two important issues in TEWS, i.e., prediction accuracy and computation time which significantly influents the first warning release time. There have been many researches conducted to predict tsunami, but the prediction accuracy and computation time achieved in those researches are not optimal yet. The objective of this research is to increase the performance of tsunami potential prediction in terms of accuracy within acceptable range of time required by the tsunami early warning system using a new methodology approach, i.e., sensor fusion and machine learning approach. The dissertation research is based on the initial hypothesis that the use of a sensor fusion approach in providing machine learning input in the form of combining seismic features obtained from seismometer data and global positioning system (GPS) data is expected to optimize tsunami prediction performance in terms of accuracy with acceptable computational time for tsunami early warning purposes. Seismic features obtained from the seismometer data include the rupture duration, the dominant period of the P signal, the period-duration discriminant, the moment magnitude, and information related to the position of the seismometer station. The seismic features obtained from GPS data are in the form of information related to focal mechanisms such as the direction of the fault's slip. This research is conducted in four stages. The first stage of the research focused on two substantial tasks i.e., designing features set model to generate a proper data set from seismic signals waveform for training and testing purposes, and doing features extraction process to result relevan seismic features. The second stage of the research focused on doing experiments and analyzing the performance of some machine learning classifier i.e., k-nearest neighbor (KNN), decision tree, random forest, support vector machine (SVM), and artificial neural network (ANN) for predicting tsunami potential based on the seismic features. In the third stage, an inversion model of Okada was developed using Genetic Algorithm (GA) to estimate some source parameters and momen magnitude of an earthquake. The inversion model is also used as features extraction process using GPS time series data. In the last stage, the research explored some sensor fusion models which involved some seismometer sensors and GPS sensors data to optimize the performance of the classifier model used in predicting the tsunami potential. Several original and novel contributions resulting from this research include a specific seismic feature set model for the purpose of predicting tsunami potential, the Okada-AG inversion model to estimate earthquake source parameters and moment magnitude, and a sensor fusion model involving seismometer sensors and GPS sensors to improve the prediction performance. At the end of the study, the performance of tsunami potential prediction has been increased and reached accuracy of more than 95% with acceptable computation time in accordance with the needs of an early warning system.