Infilling missing precipitation records using variants of spatial interpolation and data-driven methods: use of optimal weighting parameters and nearest neighbour-based corrections

Variants of spatial interpolation and data-driven methods to fill gaps in daily precipitation records are developed and evaluated in this study. The evaluated methods include variations of inverse distance and correlation weighting procedures, linear weight optimization and artificial neural networks. An already existing method, support vector logistic regression-based copula, is also assessed. Optimal weights are estimated using inverse distance and correlation-based weighting methods, post-corrections of spatially interpolated estimates for rain or no rain classifications using support vector machine (SVM), and variations of a single best classifier (SBC) are used. The optimal number of gauges for use in spatial interpolation methods and for artificial neural network-based method are selected. Three benchmark methods provide a basis against which all the methods are compared: single best estimator (SBE), and spatial and climatological mean estimators (SME and CME). All of the methods are tested for estimating varying amounts of missing precipitation data at 53 rain gauges located in South Florida, USA. Results show that the linear weight optimization method with an SBE provides the best estimates of daily precipitation values based on several performance metrics. Results from evaluation of different methods and their variants indicate use of optimized exponents in distance and correlation-based weighting methods, classifiers for rain or no rain conditions, and an optimal number of neighbours in spatial interpolation improve estimates of missing data. Corrections to missing data estimates using nearest neighbours can help in improving the accuracy of rain and no rain state determinations with a possibility of introducing bias in estimates.