Spatial interpolation of precipitation indexes in Sierra Nevada (Spain): comparing the performance of some interpolation methods

The objective of this paper is to examine the spatial distribution of several precipitation indexes in Sierra Nevada, Spain: mean annual number of wet days (R ≥ 1 mm), mean annual number of heavy rainy days (R ≥ 10 mm) and mean annual number of very heavy precipitation days (R ≥ 20 mm) and test the performance of several interpolation methods using these variables. In total, 17 univariate and multivariate methods were tested. A set of 36 metereological stations distributed in Sierra Nevada and neighbouring areas was analysed in this study. The original data did not followed the normal distribution; thus, a logarithm was applied to data meet normality purposes. Interpolator’s performance was assessed using the root mean square error generated from cross-validation. The results showed that the mean annual R ≥ 10 mm and R ≥ 20 mm have a higher variability than R ≥ 1 mm. While the elevation and longitude did not show a significant correlation with the studied indexes, the latitude (i.e. distance to the sea) showed a significant negative correlation. The regressions carried out confirmed that elevation was the covariate with higher capacity to explain the variability of the indexes. The incorporation of elevation and longitude slightly increased the explanation capacity of the models. The data of LogR ≥ 1 mm, LogR ≥ 10 mm and LogR ≥ 20 mm displayed a clustered pattern, especially the last two indexes that also showed a strong spatial dependency attributed to the effects of local topography, slope, aspect and valley orientation. The best fitted variogram model to LogR ≥ 1 mm was the linear one while for the LogR ≥ 10 mm and LogR ≥ 20 mm, the Gaussian was the most appropriate. The best interpolator for LogR ≥ 1 mm was the local polinomyal with the power of 1, whereas for LogR ≥ 10 mm and LogR ≥ 20 mm, regression kriging (ROK) using as auxiliary variables the elevation, latitude and longitude was the most accurate. ROK methods significantly improved the interpolations accuracy, especially in LogR ≥ 10 mm and LogR ≥ 20 mm. Nevertheless, the covariates, when used as auxiliary information in ordinary kriging, did not improve the precision of the interpolation.