Abstract: An estimation algorithm based on Kalman filer was developed to remove noise of the fiber temperature sensing system and to estimate the state in real time. The fiber temperature sensing system was the linear and dynamical system. Temperature was the time-discrete state variable which was modeled by Gauss-Markov random process with the additive and white Gaussian state noise. Based on Bayesian-MAP inference and MMSE criterion, the posterior state can be estimated by update equations with new measure. Given the initial parameters, the optimal estimator of temperature was achieved by such iterative process. FGC-LR, the distributed optical fiber Raman temperature sensing system, was the experimental setup with 2 km sensing fiber. The interval between adjacent sample points was 1 m. Local variance and SNR were used to evaluate the algorithm's performance in noise removal and estimation. The local variance is reduced by 83.56% at low temperature point and 84.09% at high temperature point. SNR at the normal temperature point (at 1 000 thm) and the heated points (at 1850 thm) are increased 18.45% and 16.80% respectively. These mean that the algorithm works well in noise removal of fiber sensing signal both at room-temperature and at heated points.