Improving estimation of missing values in daily precipitation series by a probability density function-preserving approach

This work presents a novel method for estimating missing values in daily precipitation series. It is aimed at identifying the event time location with good accuracy and reconstructing the correct amount of daily rainfall. In addition, the statistical properties of the time series, i.e. both probability distribution and long-term statistics, are preserved. The completion method is based on a two-step algorithm that uses information from a cluster of neighboring stations. First, wet and dry days are tagged, and subsequently, the full precipitation amount for wet-classified days is estimated by a modified multi-linear regression approach. This method avoids overestimation of the number of wet days and underestimation of intense precipitation events, which are typical side effects of common regression-based approaches.