Applying the Particle Filter to the Volcanic Ash Tracking PUFF Model for Assimilating Multi-Parameter Radar Observation

Abstract

In this study, a new data assimilation method called “particle filter” was applied to the volcanic plume tracking model called PUFF to assimilate the Multi-parameter (MP) radar observations at Sakura-jima volcano. In the particle filter algorithm, the statistical likelihood was computed for each ash particle of the model given the observed MP radar data. Particles with a high likelihood were retained, but particles with a low likelihood were removed from the computation. The removal was followed by resampling of new particles at high-likelihood locations. The results show that the particle filter works properly to generate suitable new particles in the open space between the model and the observed particles. As the plume shape of MP radar observation is an important information source, observed particles were added at the resampling stage of the particle filter. A proper threshold value for removing or retaining the particles was examined using likelihood estimation. In this study, we determined the proper threshold to resample approximately half of the model particles. The results of the analysis show a reasonable mix of observed, predicted, and resampled data at proper locations, filling the open space between the prediction and the observation. It was found that data assimilation using the particle filter is a suitable method for assimilating the MP radar observation to the PUFF model prediction. This study demonstrates that the new PUFF model system, combined with real-time MP radar data using a particle filter, is highly reliable and useful for preventing volcanic hazards around active volcanoes.