Estimation of Missing Daily Temperatures: Can a Weather Categorization Improve Its Accuracy?

Abstract

A method of estimating missing daily temperatures is proposed. The procedure is based on a weather classification consisting of two steps: principal component analysis and cluster analysis. At each time of observation (0700, 1400, and 2100 local time) the weather is characterized by temperature, relative humidity, wind speed, and cloudiness. The coefficients of regression equations, enabling the missing temperatures to be determined from the known temperatures at nearby stations, are computed within each weather class. The influence of various parameters (input variables, number of weather classes, number of principal components, their rotation, type of regression equation) on the accuracy of estimated temperatures is discussed. The method yields better results than ordinary regression methods that do not utilize a weather classification. An examination of statistical properties of the estimated temperatures confirms the applicability of the completed temperature series in climate studies.