An assessment of statistical interpolation methods suited for gridded rainfall datasets

Accurate spatial distribution information of rainfall is essential to rainfall-induced hazard predictions and statistical interpolation methods may serve as a useful tool to produce a detailed distribution from coarse data sources. Although numerous comparison studies about different interpolation methods have been conducted on irregular rain-gauge networks, there is a need to perform such work on the increasingly available gridded rainfall data. Carried out in the Emilia-Romagna region (Italy) from 2008 to 2018, this study aims to examine accurate and appropriate interpolation methods to produce finer rainfall surface maps based on the 0.25° × 0.25° ERA5 gridded precipitation datasets. Five interpolation techniques, namely Thiessen polygons, Inverse Distance Weighting (IDW), Thin Plate Spline (TPS), Ordinary Kriging (OK), and ordinary Co-Kriging (CoK), have been selected and compared at different time scales (annual, monthly, and annual maximum daily precipitation). To assess the accuracy, the leave-one-out-cross-validation（LOOCV）test was used by using the indexes of Bias, Correlation coefficient (CC), the Nash-Sutcliffe Efficiency (NSE), Root Mean Square Error (RMSE) and the Kling-Gupta Efficiency (KGE). Additionally, visual inspections are employed to evaluate the plausibility of the interpolated maps. Results show: (1) All five interpolation methods have certain capabilities to improve spatial resolution, but they fail on accuracy at the daily scale. The OK generally outperforms the other four methods, while TPS shows better performance through the visual inspection at the monthly scale; (2) Unlike in the case of the interpolation using conventional point-based rain-gauge data, the multivariate method CoK is inferior to the univariate ones and the p-power of the IDW also differs; (3) Winter and spring results are better than those of summer and autumn. This study has provided useful guidance on choosing suitable rainfall interpolation methods for gridded datasets, which can be expanded in other regions and data sources to explore the generality of the conclusions.