Flood Water Level Prediction and Tracking Using Particle Filter Algorithm Fazlina Ahmat Ruslan #1 , Zainazlan Md Zain #2 , Ramli Adnan #4 # Faculty of Electrical Engineering Universiti Teknologi MARA 40450, Shah Alam, Selangor, Malaysia 1 fazlina419@salam.uitm.edu.my 2 hajizainazlan@salam.uitm.edu.my 4 ramli324@salam.uitm.edu.my Abd Manan Samad *3 * Dept. of Surveying Science and Geomatics, Faculty of Arc., Planning and Surveying Universiti Teknologi MARA 40450, Shah Alam, Selangor, Malaysia 3 dr_abdmanansamad@ieee.org Abstract— Most of the countries have paid great attention to flood water level monitoring and tracking because flood may damages people’s life and property. Since flood water level fluctuate highly nonlinear, it is very difficult to predict the flood water level. The particle filter algorithm is well known as a very effective solution for handling nonlinear problems. Thus, in this paper, this algorithm is applied to predict the flood water level. There are many variations of particle filter. This paper proposes Sequential Importance Sampling (SIS) particle filter to solve the above mentioned problem. SIS is the basic particle filter. However, the problems with SIS particle filter are the particle degeneration phenomenon, when after a few iterations only a few particles have nonzero weight. So, Sampling Importance Resampling (SIR) particle filter is also introduced as the improved particle filter. From the simulation results using Matlab, SIR particle filter outperforms SIS particle filter by comparing the Root Mean Square Error (RMSE) value. Keywords — Water Level, Sequential Importance Sampling (SIS), Sampling Importance Resampling (SIR), Particle Degeneration, Root Mean Square Error (RMSE). I. INTRODUCTION Flood disaster always causes life and economic loss. Hence, 40% of total economic loss is due to flood disaster [1]. As an enormous threat to human life and property, flood disaster always affecting large areas of the community [2]. Recently, due to rapid economic development, flood disaster has become more goaded [3]. Therefore, we have to develop a precise technique for flood water level prediction and monitoring as an alarming system to prevent future disasters. Research on flood water level prediction has long been a subject of interest. Researcher regularly generate statistical model based on past data in conventional way to predict flood water level. So they have to deal with pattern recognition [4]. To provide an alternative approach for flood water level prediction, particle filter which is capable for estimation and tracking of nonlinear and dynamic systems is presented [5]. Particle filter have been successfully applied in state estimation problem such as visual tracking [6], speech recognition [7], mobile robot localization [8], navigation [9] and fault detection [10] over the past years. The key advantage of particle filter is they can solve non-Gaussian and nonlinear state estimation problem because of their ability to represent arbitrary probability densities. This paper was organized in the following manner: Section II describes the theory of particle filter algorithm; Section III describes methodology; Section IV is on results and discussion; and finally, Section V is the conclusions. II. PARTICLE FILTER A. A. Theory of Particle Filter Based on Monte Carlo simulation, particle filter is one type of recursive Bayesian filter. Often called as bootstrap filter [11], particle filter also have similarities with particle system approximation [12], Monte Carlo filter [13], likelihood weighting algorithm [14] and etc. The particle filter algorithms start by partitioning the state space in various parts. Then, based on probability measure, the particles are filled accordingly. The concentration of the particles totally depends on the probability measure. As the probability measure increases, the concentration of the particles also increases. By using Fokker Planck Kolmogorov (FPK) equation [15], the concept of probability density function (pdf) is explained. As for state equation, it explained the particle filter algorithm evolving with time. Number of particles that represents the evolving pdf can be obtained by point mass histogram and random sampling of state space. However, we would prefer to choose another distribution because the posterior density model is difficult to sample. In Bayesian statistics, in order to avoid intractable integration, the posterior distribution or density is empirically represented by a weighted sum of number of particles, p N samples drawn from the posterior distribution; 2012 IEEE 8th International Colloquium on Signal Processing and its Applications 978-1-4673-0961-5/12/$31.00 ©2012 IEEE 431