Is it necessarily a normally distributed data for kriging? A case study: soil salinity map of Ghahab area, central Iran

After collecting data, in researches, the type of data distribution must be determined; since any analysis requires the distribution of its own data. Soil properties, such as salinity, are also the same case. Due to its direct and indirect effects on plant growth, soil salinity is an important feature that has always been investigated in agriculture and natural resources, leading to a lot of researches. These researches have often focused on the mapping of salinity using different interpolation methods and their accuracy. But the effect of the data distribution on the analysis process has been less considered. Accordingly, the purpose of this study is to investigate the effect of the distribution of soil salinity data on soil salinity mapping using Kriging method. For this purpose, 610 soil samples were taken from 0-50 cm soil depth based on a grid method and their salinity (Electrical Conductivity, EC) was determined in saturated paste extracts. Variography operations for data were performed based on both, the original distribution of the data and the usual data distribution employed for Kriging i.e. normal distribution. Salinity maps were obtained for both data distributions. Estimations were made using cross-validation approach. According to the findings, only the fitness criterion (R2) is not enough to select the optimal variogram model, while other criteria such as the proportion of the spatial structure, residual sum of square (RSS) and the nugget effect should be analyzed as well. The results showed: 1- the accuracy of the estimation based on the original distribution of the data, (i.e. non-normal distribution) which was greater than the accuracy of the estimated data using normal distribution; 2- the predictions and errors from the both, normally and non-normally distributed data did not have the normal distribution and 3 - data transformation had no effect on the normalization of the distribution of the predictions and the errors. Therefore, it is suggested that in the Kriging, in addition to the conventional method, i.e., performing Kriging using normal distribution data, the original data with non-normal distribution should also be analyzed. Finally, the type of data distribution and the optimal variogram model could be selected by comparing the obtained results.