Using the nonparametric k-nearest neighbor approach for predicting cation exchange capacity

The objectives of this study were to apply a k-NN approach to predict CEC in Iranian soils and compare this approach with the popular artificial neural network model (ANN). In this study, a data set of 3420 soil samples from different parts of Iran was used. Two different sets of cheaper-to-measure soil attributes were selected as potential predictors. The first set consisted of clay, silt, sand and organic carbon (OC) contents. The second data set was constructed using OC and clay contents. Two 'design-parameter' parameters should be optimized before application of the k-NN approach. Results showed that the algorithm efficiency is not dependent on these parameters. A wide range of suboptimal values around the optimal values may cause a slight error in terms of estimation accuracy. However, the optimal settings of the design-parameters depend on the size of the development/reference data set. In both k-NN and ANN models, the higher number of input variables can relatively improve the estimation of CEC. But this improvement was not statistically significant at the 0.05 level. Furthermore, the results showed that increasing the size of the reference data set to a certain amount (N = 1200) reduced the estimation error significantly in terms of root-mean-squared residuals (RMSE). However, no significant difference in the accuracy of k-NN and ANN methods was detected in the reference data set sizes for N>1200. Results showed no significant difference between this approach and ANN models, suggesting the competitive advantage of the k-NN technique over other techniques to develop pedotransfer functions (PTFs), for example, the redevelopment of PTFs is not necessarily required as new data become available. (C) 2015 Elsevier B.V. All rights reserved.