Comparison of missing value imputation methods for Turkish monthly total precipitation data

Abstract

Climate related studies are required complete time series data to be used. On the other hand, considerable number of observations is missing in meteo- rological time series due to several reasons. This con°icting problem, however, can be overcame by imputing missing values using observations of correlated nearby climate stations. The main aim of this study, therefore, is to compare the performances of six di®erent methods for imputing monthly total precipi- tation series obtained from stations located in two di®erent climate regions of TДurkiye. These include Single Arithmetic Average (SAA), Normal Ratio (NR), NR Weighted with Correlations (NRWC), Multi Layer Perceptron type Neural Network (MLPNN) and Expectation-Maximization Algorithm based on Monte Carlo Markov Chain (EMMCMC). In addition, we propose a modiЇcation in the EMMCMC method which uses the results of di®erent imputation methods as reference series. Results show that both EMMCMC methods perform better than the other imputation methods considered in the study.