Data assimilation with an improved particle filter and its application in the TRIGRS landslide model

Abstract

Abstract. Particle filters have become a popular algorithm in data assimilation for their ability to handle nonlinear or non-Gaussian state-space models, but they have significant disadvantages. In this work, an improved particle filter algorithm is proposed. To overcome the particle degeneration and improve particles' efficiency, the processes of particle resampling and particle transfer are updated. In this improved algorithm, particle propagation and the resampling method are ameliorated. The new particle filter is applied to the Lorenz-63 model, and its feasibility and effectiveness are verified using only 20 particles. The root-mean-square difference (RMSD) of estimations converges to stable when there are more than 20 particles. Finally, we choose a peristaltic landslide model and carry out an assimilation experiment of 20 days. Results show that the estimations of states can effectively correct the running offset of the model and the RMSD is convergent after 3 days of assimilation.