Evaluation of imputation techniques for infilling missing daily rainfall records on river basins in Ghana

Abstract

Statistical imputation techniques were evaluated for infilling missing records in daily rainfall data within the Pra and Densu river basins in Ghana. The imputation techniques considered were mean, regression, multiple imputation by chain equation, k-nearest neighbour, probabilistic principal component analysis (PPCA), missForest, linear interpolation, hot deck, expectation maximization, Gaussian copula, inverse distance weighting and kriging. Different percentages of missing records (5%, 10%, 20% and 30%) were artificially introduced into the complete datasets. Then, the missing records were imputed and compared with the observed values. The root mean square error, mean absolute error, bias, coefficient of determination, similarity index and Kolmogorov-Smirnov performance statistics were used to evaluate the methods. The results were mixed depending on the performance metric used. However, the best candidates were regression, PPCA and missForest imputation techniques. These methods were better for estimating the numbers of dry and wet periods and the moderate to extreme rainfall values.