Comparing interpolation methods to predict soil total phosphorus in the Mollisol area of Northeast China

Abstract

Soil total phosphorus (TP) is an essential indicator to reflect the soil fertility in agricultural ecosystems. The accurate prediction of the spatial heterogeneity of TP is crucial to evaluate the soil productivity and quality. In this study, the interpolation methods of inverse distance weighted (IDW), radial basis functions (RBF), ordinary kriging (OK), co-kriging (COK), multiple linear regression (MLR), geographically weighted regression (GWR), regression kriging (RK), and geographically weighted regression kriging (GWRK) were used to estimate the spatial patterns of TP in four Mollisol areas with different kinds of landscapes at diverse scales. The calculation method, accuracy (mean error, mean absolute error, root mean square error and relative improvement), cost (money, labour and time), and process by which the models were established were also compared in order to ascertain the best interpolation methods to analyse and predict the spatial heterogeneity of TP in Mollisol areas. The results showed that i) generally, compared with OK, the RK and GWRK methods typically increased the simulative accuracy of the spatial distribution of TP, while IDW, COK and MLR decreased the accuracy, and RBF and GWR were not consistent at improving the accuracy in the Mollisol area; ii) the slope steepness and brightness index could be introduced to regression models as superior auxiliary variables to improve the interpolation precision (0.8%–1.9%) of TP by RK and GWRK when the study area is relatively flat; iii) the GWRK and RK methods with smaller root mean square error and higher relative improvement outperformed other methods in the region with suitable sampling evenness (>45%), while the RBF method is an optional approach when the sampling evenness is low. In summary, the GWRK and RK could be regarded as the optimal interpolation methods, despite the fact that their improvement was limited in this study. Meanwhile, when considering the cost, time and process by which the models were established, OK can also be deemed an optional interpolation method with relatively acceptable accuracy. Furthermore, the evenness and density of sample should be considered when mapping soil TP in Mollisol areas.