Symmetrical uncertainty and random forest for the evaluation of gridded precipitation and temperature data

Abstract

Selection of appropriate gridded rainfall and temperature data is a key problem for hydro-climatic studies, particularly in regions where long-term reliable and dense observations are not available. The ability of two intelligent algorithms, symmetrical uncertainty (SU) and random forest (RF), to assess the degree of similarity or the distance between two time series was utilized in this study for the evaluation of gridded climate data. In this study, the performances of seven widely used gridded rainfall datasets and five temperature datasets were evaluated against the available station data in Egypt. Monthly rainfall and mean temperature data recorded at 57 locations for the period 1979–2014 were used for this purpose. The results revealed the better performance of Global Precipitation Climatology Centre (GPCC) gridded rainfall and University of Delaware (Udel) gridded temperature data in replicating observed rainfall and mean temperature, respectively, in most of the locations in Egypt. Validation of the results using conventional statistical metrics revealed the better performance of different datasets in term of different metrics at different locations. However, the mean values of all the metrics support the results obtained using SU and RF. The study indicates that SU and RF can be used for the selection of appropriate gridded rainfall and temperature data by avoiding confusion arise from contradictory results obtained using various statistical metrics.