Using Residual Resampling and Sensitivity Analysis to Improve Particle Filter Data Assimilation Accuracy

Abstract

Data assimilation (DA), an effective approach to merge dynamic model and observations to improve states estimation accuracy, has been a hot topic in the earth science and lots of efforts have been devoted to the DA algorithms. In this paper, an improved residual resampling particle filtering (improved RR-PF) is proposed. Compared with the generic residual resampling particle filtering (generic RR-PF), the improved RR-PF not only solves the degradation of particles, but also maintains the diversity of particles. Besides, sensitivity analysis is carried out to analyze the impact of some parameters to assimilation and to determine the optimal parameters. These parameters are of significant importance to DA but cannot be determined easily. Finally, soil moisture from Soil Moisture Experiment 2003 and VIC model simulations were assimilated with the improved RR-PF with parameters determined by the sensitivity analysis. The result shows that the accuracy of soil moisture greatly improves after DA. Compared with generic RR-PF, the performance of improved RR-PF is superior in accuracy and diversity of particles.