Abstract
Precipitation is characterized by variability in space and time. In addition, there are many factors affecting the magnitude and distribution of precipitation, such as altitude, various air mass movements, distance from the moisture sources, temperature, pressure, and topography. The magnitude and distribution of precipitation vary temporally and spatially even in small areas. However, the precipitation predictor should not be fixed with time and space, but adapt itself to the evolving meteorological conditions. Describing and predicting the precipitation variability in space and/or time are fundamental requirements for a wide variety of human activities and water project designs. Forecasting models can be classified into two categories, those models that have fixed parameters and variances, and likewise another group of models with varying parameters and variances. Models with fixed parameters require stationarity in both the mean and variance throughout the entire range of observations. That is why so much effort is spent to make the data stationary in the mean and variance. Otherwise, the results are not meaningful from a statistical point of view. For example, when the data pattern changes as with a step or trend, or when there are transient shifts, classical statistical theory will treat those as random effects or temporary shifts. If the changes are continuous, a new forecasting model will have to be specified to deal with the new equilibrium conditions. However, the model will be good for those new equilibrium conditions only when fixed patterns exist. Kalman filter (KF), can deal with step changes and transient situations because they update their parameters in a way that takes account of changes in pattern. The objective of this thesis is to investigate and develop a KF model approach to multisite precipitation modeling. In order to see the effectiveness of the KF model developed in this thesis, 30 year records (1956-1985) of annual rainfall for the 52 different meteorology stations are used, and these stations are distributed approximately covering all of Turkey with more concentration in the northwestern part. The necessary contour maps of observed and estimated precipitation amounts are attuned through the results of the software developed during the course of this study. Furthermore, regional error distribution maps are also attuned for any year. The results indicates that KF provide an efficient method for modelling annual rainfall in both time and space dimensions.